online social networks and media - cnr · book 1 or a “retweet” on twitter 2 ). in the first...
TRANSCRIPT
Online Social Networks and Media 3–4 (2017) 75–97
Contents lists available at ScienceDirect
Online Social Networks and Media
journal homepage: www.elsevier.com/locate/osnem
Recommender Systems for Online and Mobile Social Networks: A
survey
Mattia G. Campana
∗, Franca Delmastro
IIT-CNR, Via G. Moruzzi 1, 56124 Pisa, Italy
a r t i c l e i n f o
Article history:
Received 24 September 2017
Revised 28 October 2017
Accepted 30 October 2017
Keywords:
Recommender Systems
Online Social Networks
Mobile Social Networks
a b s t r a c t
Recommender Systems (RS) currently represent a fundamental tool in online services, especially with the
advent of Online Social Networks (OSN). In this case, users generate huge amounts of contents and they
can be quickly overloaded by useless information. At the same time, social media represent an important
source of information to characterize contents and users’ interests. RS can exploit this information to fur-
ther personalize suggestions and improve the recommendation process. In this paper we present a survey
of Recommender Systems designed and implemented for Online and Mobile Social Networks, highlighting
how the use of social context information improves the recommendation task, and how standard algo-
rithms must be enhanced and optimized to run in a fully distributed environment, as opportunistic net-
works. We describe advantages and drawbacks of these systems in terms of algorithms, target domains,
evaluation metrics and performance evaluations. Eventually, we present some open research challenges
in this area.
© 2017 Elsevier B.V. All rights reserved.
Contents
1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
2. The recommendation task . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
2.1. Evaluation metrics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
2.1.1. Prediction accuracy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
2.1.2. Ranking accuracy. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
3. Main approaches for Recommender Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
3.1. Collaborative Filtering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
3.1.1. Memory-based . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
3.1.2. Model-based . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
3.2. Content-based Recommender Systems. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
3.3. Graph-based . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
3.4. Context-Aware Recommender Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80
4. Recommender systems for Online Social Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
4.1. Social-aware recommendations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
4.1.1. Friends and people recommendations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
4.1.2. Items recommendations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
4.2. Tag-based recommendations. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
4.3. Location-aware recommendations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
5. Recommender Systems for Mobile Social Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
5.1. Collaborative Filtering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
5.1.1. OppCF . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
5.1.2. MobHinter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90
∗ Corresponding author.
E-mail addresses: [email protected] (M.G. Campana), [email protected] (F. Delmastro).
https://doi.org/10.1016/j.osnem.2017.10.005
2468-6964/© 2017 Elsevier B.V. All rights reserved.
76 M.G. Campana, F. Delmastro / Online Social Networks and Media 3–4 (2017) 75–97
5.1.3. DiffeRS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90
5.1.4. LocPref . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90
5.2. Exploiting tags . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
5.2.1. Information-Centric Habit (ICe-Habit) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
5.2.2. Pervasive PLIERS (p-PLIERS) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
6. Concluding remarks and open challenges . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92
Acknowledgments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
Fig. 1. Recommender System architecture.
f
a
t
t
s
s
a
a
l
o
D
d
p
o
p
T
t
l
a
r
v
r
t
i
a
l
m
2
e
d
t
f
i
m
f
1. Introduction
Recommender Systems (RS) and Online Social Networks (OSN)
have established a strong cooperation in the last few years. They
both aim at coping with the huge amount of data produced and
shared by users through online platforms, trying to maintain a
high user engagement. This cooperation is built upon the ad-
vantages that both systems can achieve: the optimization of the
recommendation techniques, by exploiting additional content and
user characterization derived from OSN, and the increasing request
of OSN services’ personalization. RS targeting the social media do-
main have been already defined in the literature as Social Recom-
mender Systems (SRS) [1,2] . However, this notion should be further
expanded, embracing also the recent evolution of social media to-
wards the pervasive and mobile computing environment. Here, the
network can be also implemented by the physical co-location of
users and devices, and their ability to exploit direct wireless com-
munications for content sharing and dissemination, without re-
quiring a constant Internet access. In this direction, the concept
of Mobile Social Networks (MSN) has been introduced in [3] , not
as a simple extension of OSN services running on mobile devices,
but as the opportunity to create real networks of people , in which
users actively participate in the generation and sharing of contents,
anywhere and anytime, based on opportunistic networking and
device-to-device (D2D) communications. In this scenario, RS can be
optimized by exploiting additional information derived from users’
mobile devices (i.e., sensors), which help contextualize the user’s
preferences, and to rely on a partial and dynamic knowledge of
the network and the available data. This concept contributes to the
migration towards a new Internet paradigm, the Internet of People
(IoP) [4] , in which users are at the center, by actively contributing
to the network evolution both from the communication and data
point of views.
In this paper we present a survey of RS defined both for OSN
and MSN, highlighting advantages and drawbacks of standard rec-
ommendation techniques applied in these environments, and how
those systems evolved over time, in terms of technical solutions,
target domains, evaluation metrics and performance evaluations.
In Section 2, we summarize the problem of recommendations and
the evaluation metrics generally used in the literature. Then, in
Section 3 , we present an overview of standard techniques like Col-
laborative Filtering (CF), Content-based RS, Network-based RS and
Context-aware RS, with particular attention to the various infor-
mation used to characterize the relationship between users and
items to further optimize and personalize the recommendations.
This section would also introduce the reader with standard nota-
tions and methods applied in RS, which will be analyzed in the
subsequent sections. In Sections 4 and 5 , we describe the main so-
lutions presented in the literature for OSN and MSN, respectively.
The first area has been widely studied in the last years, proposing
solutions that can address different target domains (e.g., to recom-
mend people, locations, point of interests, tags or contents) by ex-
ploiting heterogeneous context information. Therefore, we decided
to group the proposed solutions by the type of context information
used (e.g., social relationships, tags, location) and the recommen-
dation target (e.g., to recommend contents, tags, friends, people).
The research area of RS for MSN is still in its infancy, and
ew solutions have been presented in the literature, mainly aimed
t optimizing content dissemination in opportunistic networks
hrough personalized recommendations. The main difference be-
ween RS for OSN and MSN relies on the knowledge they use to
elect the recommendations for their final users. In OSN, RS as-
ume to have access to a complete knowledge of all the avail-
ble objects in the network (i.e., contents, tags, etc.), residing on
centralized infrastructure. Instead, in MSN, RS can rely on the
ocal knowledge of each user, represented by the local available
bjects and those declared by other users in proximity through
2D communications. In this scenario, each mobile device has a
ifferent knowledge of the network, which grows up through its
ersonal mobility and its opportunities to communicate with the
thers. In addition, mobile devices have limited computational ca-
abilities and RS must be characterized by efficient response time.
his is due to the limited and unpredictable duration of the oppor-
unistic contact, during which mobile devices can exchange their
ocal knowledge and the recommended objects. In order to evalu-
te and compare the performance of RS for MSN, it is necessary to
eproduce the realistic behavior of mobile users in a synthetic en-
ironment, since a common evaluation framework and appropriate
eal datasets are not currently available. In Section 5 we present
he proposed RS in this area, highlighting the advantages of us-
ng context information to improve the recommendation process
nd demonstrating their efficiency with respect to centralized so-
utions. Eventually, in Section 6 we present some concluding re-
arks and open research challenges in this area.
. The recommendation task
As a general concept, RS try to identify and foresee users’ inter-
sts in specific contents, based on their previous experiences. Fig. 1
epicts the high-level architecture of a typical RS. When a user in-
eracts with the system, she provides a set of explicit or implicit
eedbacks (e.g., likes, clicks, ratings) about her tastes. For example,
f a user positively rates an article about a new smartphone, she
ay also be interested in reading news about mobile apps. There-
ore, the basic idea of RS is to exploit this information to infer user
M.G. Campana, F. Delmastro / Online Social Networks and Media 3–4 (2017) 75–97 77
i
p
i
t
g
s
t
R
R
w
i
r
s
p
d
c
l
b
c
t
a
i
(
i
b
a
o
o
s
r
e
s
a
l
t
t
p
i
t
e
t
m
m
t
b
2
p
c
v
p
i
i
p
t
b
w
R
i
t
n
a
o
2
i
t
b
t
r
p
a
a
o
E
t
e
o
N
f
R
M
w
r
v
t
m
p
2
r
l
m
m
e
(
G
D
I
p
r
a
t
t
t
nterests. Based on the user’s past feedbacks, RS learn a model to
redict how much a user can be interested in new items. Those
tems are then ranked according to their predicted relevance for
he user. Finally, the higher ranked items will be proactively sug-
ested to the user.
The relationship between users and items is generally repre-
ented by a matrix R M
storing a rating r u, i to identify how much
he user u liked the item i in the past. More formally, the matrix
M
is called ratings matrix and it can be defined as follows:
M
: U × I → R, (1)
here U = { u 1 , . . . , u m
} represents the set of users, I = { i 1 , . . . , i n }s the set of items, and R = { r 1 , . . . , r k } denotes the set of possible
atings that users can use to express their grade of interest for a
pecific item.
The semantic value of the ratings highly depends on the ap-
lication domain. They are often specified as a discrete set of or-
ered numbers (e.g., a 5-point rating scale is commonly used in e-
ommerce websites, like Amazon, and video-on-demand services,
ike Netflix), or a binary value, as in most OSN (i.e., “like” on Face-
ook 1 or a “retweet” on Twitter 2 ). In the first case, the ratings are
onsidered as explicit feedbacks of the users, while in the latter
hey are considered as implicit feedbacks , derived from the users’
ctivities in OSN. In this paper, we generally refer to ratings scale
ncluding both cases.
A missing value in the ratings matrix can have two meanings:
i) the user does not want to express an opinion about a specific
tem, or (ii) the user does not know that item yet.
As we will discuss in details in the next sections, a great num-
er of recommendation techniques (e.g., Collaborative Filtering)
ims to predict missing ratings in R M
in order to generate lists
f new items to recommend to the users. However, there are also
ther Recommender Systems (e.g., some network-based RS) that
imply provide ranked lists of items without explicitly predicting
atings, using the ratings matrix just to model the users’ prefer-
nces.
Typically, there are a lot of missing values in R M
. In [5] the
parsity of the ratings matrix has been calculated to be gener-
lly larger than 99% in commercial systems. This is due to the
ong tail property [6] , which is satisfied by different real-world set-
ings. According to this property, just a small subset of the con-
ents in the system is associated with a high number of rates (i.e.,
opular items ) and, consequently, the vast majority of the items
s rarely rated by the users. This characteristic of the ratings ma-
rix has an important implication for RS: the system may not have
nough information to generate relevant recommendations for the
arget user, and the presence of popular items may bias the recom-
endation process, likely providing trivial recommendations. The
ain recommendation approaches presented in Section 3 address
his problem in different ways. However, before detailing them, we
riefly describe the most used evaluation metrics.
.1. Evaluation metrics
RS performance can be measured in terms of accuracy in
redicting ratings or improvement of the user’s experience. This
hoice depends on the RS target: to increase the profit of the ser-
ice or to improve the user’s satisfaction.
In addition, RS can be evaluated through online or offline ap-
roaches. The online approach mainly leverages on the continuous
nteraction between the users and the system; the user selects an
tem among those recommended, and this choice is used as in-
ut to machine learning algorithms used to adapt the RS behavior
1 https://www.facebook.com
2 https://twitter.com
f
r
i
o the user’s preferences. This approach requires that a large num-
er of users is involved in the RS evaluation, and only few research
orks explored this kind of evaluation until now (e.g., [7–9] ). Most
S are currently evaluated using the offline approach by exploit-
ng available datasets including the history of the users actions. In
hese cases, the datasets are generally divided into two subsets,
amely the training ( T r ) and test ( T e ) sets. The former represents
set of examples used to fit the parameters of the proposed rec-
mmendation model, and the latter is used to test its accuracy.
.1.1. Prediction accuracy
The prediction accuracy measures how much the predicted rat-
ngs, derived from the learning phase of the RS, differs from the ac-
ual ratings computed on the test set T e . More formally, let r jz ∈ T e e the explicit rating given by the user u j for the item i z in the
est set, and ˆ r jz be the rating predicted by the RS for the user u j elated to the item i z , the accuracy is determined by the error com-
uted for the user–item couple ( u j , i z ) ∈ T e as e jz = r jz − r jz . Lever-
ging on the single e jz errors for each pair < user, item > , different
ccuracy metrics have been proposed to compute the overall error
ver the entire test set T e . The most used are Root Mean Squared
rror ( RMSE ) [10] and Mean Absolute Error ( MAE ) [10] . However,
hey both depend on the specific rating scale used in the refer-
nce dataset. Thus, in order to compare the performance of a RS
ver different datasets, we should use their normalized versions:
ormalized RMSE ( NRMSE ) and Normalized MAE ( NMAE ), defined as
ollows:
M SE =
√ ∑
(u j ,i z ) ∈ T e e 2 jz
| T e | ; NRM SE =
RM SE
r max − r min
, (2)
AE =
∑
(u j ,i z ) ∈ T e ∣∣e jz ∣∣
| T e | ; NM AE =
M AE
r max − r min
, (3)
here r max and r min are, respectively, the maximum and minimum
atings in the dataset. Both RMSE and MAE (and their normalized
ersions) are meaningful accuracy measures: smaller values, bet-
er RS performance. However, as RMSE sums the squared errors, it
ay be highly affected by outliers. This means that a few wrong
redictions could lead to a worst overall RMSE.
.1.2. Ranking accuracy
Ranking accuracy deals with the different levels of utility of the
ecommended items with respect to their position in the ranked
ist proposed to the user. Ideally, the best ranking in the recom-
endation list should be assigned to the items considered as the
ost useful to the user. One of the most popular metrics used to
valuate the ranking accuracy is the Discounted Cumulative Gain
DCG ) [11] and its normalization, Normalized Discounted Cumulative
ain ( NDGC ) [12] , which are defined as follows:
C G =
1
m
m ∑
u =1
∑
j∈ I u
g u j
log 2 (v j + 1) ; NDC G =
DC G
IDC G
, (4)
In DCG formula, m is the number of users in the test set T e ,
u represents the set of items rated by the user u , and v j is the
osition of the item j in the recommended list. The numerator, g uj ,
epresents the utility (i.e., the gain) of the item j for the user u ,
nd the denominator represents a discount factor with respect to
he item position in the ranking list. The utility directly derives by
he ground-truth (i.e., the actual rating r uj provided by the user in
he test set). The Ideal value IDCG is computed with the same DCG
ormula by using the ground-truth ranking.
Generally, a RS suggests to the users the top- k elements of the
ecommendation list. An alternative way to evaluate its accuracy
s to consider the trade-off between the length of the list R and
L
78 M.G. Campana, F. Delmastro / Online Social Networks and Media 3–4 (2017) 75–97
Fig. 2. Example of matrix factorization.
s
i
i
s
r
f
r
w
k
t
a
d
e
t
a
e
b
t
P
w
t
t
t
u
P
w
i
c
b
c
p
e
t
t
b
o
3
a
s
a
n
i
m
the number of actual relevant items for the user (i.e., the items
positively rated by the user in the test set).
The relevant items contained in R L are also identified as the
true-positives ( tp ), while the relevant items not included in R L are
called false-negative ( fn ). In the same way, the proposed items not
actually relevant for the target user are called false-positive ( fp ),
while those not relevant and discarded are called true-negative ( tn ).
Therefore, the trade-off between the length of R L and the number
of relevant items can be measured using the Precision and Recall
measures defined as follows [13] :
P recision =
t p
t p + f p ; Recall =
t p
t p + f n
(5)
In addition, one way to create a single metric that summarizes
both the aforementioned measures is the F-measure [14] , which is
the harmonic mean of equally weighted Precision and Recall:
F =
2 · P recision · Recall
P recision + Recall (6)
3. Main approaches for Recommender Systems
3.1. Collaborative Filtering
Collaborative Filtering (CF) is considered the most popular and
widely implemented technique in RS [15] . The underlying assump-
tion of CF is that people with similar preferences will rate the
same objects with similar ratings [16] . Existing CF solutions can be
categorized in two main classes: (i) memory-based and (ii) model-
based methods. Memory-based solutions leverage on similarities in
users’ behaviors and preferences to make inferences about missing
values in the ratings matrix. Instead, model-based methods exploit
the matrix values to learn a model, similarly to a classifier that
trains a model from labeled data. The learned model is then used
to predict the relevance of new items for the users.
3.1.1. Memory-based
Memory-based algorithms (also known as Neighborhood-based )
rely on the notion of similarity among users, or items, to predict
the possible interest of a user in items that she has not seen (or
rated) before. In the literature, memory-based CF solutions are typ-
ically divided in two main categories: user-based and item-based .
The former approach is based on the assumption that similar users
typically rate the items in a similar way. Therefore, a user-based CF
predicts the rating that a user u might assign to an item by aggre-
gating the ratings that the most similar users to u have previously
given to that item. Formally, the predicted rating of a user u to the
item j can be formulated as follows:
ˆ r u, j =
1
K
∑
k ∈ N u Sim (u, k ) · r k, j , (7)
where N u is the set of the K users most similar to the target user
u; Sim ( u, k ) represents the similarity between the users u and k
for a predefined similarity measure, and r k, j represents the rating
made by the user k for the item j .
In contrast with user-based CF, item-based CF focuses on the
imilarities among items. It is based on the assumption that similar
tems are rated in a similar way by the same user. In this case, the
tems recommended to the user u are ranked by aggregating the
imilarities between each candidate item and the items that u has
ated in the past. It is possible to formulate the prediction rating
or the item-based CF as follows:
ˆ u, j =
1
K
∑
k ∈ N i Sim ( j, k ) · r u,k , (8)
here N i represents the set of neighbor items of item j , and Sim ( j,
) is the similarity value (for a predefined similarity measure) be-
ween items j and k .
The similarity computation between users (or items) represents
crucial step in memory-based approaches, as it can seriously re-
uce both the accuracy and the performance of RS [17] . In the lit-
rature, a number of similarity measures have been proposed, but
he Pearson Correlation (PC) similarity seems to provide the most
ccurate results [18] . PC selects just the co-rated items and consid-
rs the differences in the mean and variance of the ratings made
y two users u and v . PC similarity between two users is given by
he following formula:
C(u, v ) =
�i ∈ I u, v (r ui − r u ) · (r v i − r v ) √
�i ∈ I u, v ( r ui − r u ) 2 · �i ∈ I u v ( r v i − r v ) 2 , (9)
here I u, v is the set of items rated by both users u and v, r ui is
he rating made by u for the item i , and r u is the mean value of
he ratings made by u . On the other hand, PC similarity between
wo items i and j can be calculated comparing the ratings made by
sers that have rated both items:
C(i, j) =
�u ∈ U i, j (r ui − r i ) · (r u j − r j ) √
�u ∈ U i, j ( r ui − r i ) 2 · �u ∈ U i j
( r u j − r j ) 2 , (10)
here U i, j is the set of users who rated both items i and j , and r is the average rating value obtained by the item i . For a complete
omparison of the similarity measures typically used in memory-
ased CF, we refer the reader to [17] .
The main drawback of memory-based CF is represented by the
omputational cost: the computation of similarities between all
airs of users, or items, typically requires a quadratic time. How-
ver, they gained popularity due to their very simple implementa-
ion, providing an intuitive justification for the computed predic-
ions. For instance, in item-based CF, the list of similar items can
e presented to the user in order to better understand the reason
f the recommendation.
.1.2. Model-based
Although memory-based CF approaches are easy to implement
nd useful in effectively predicting missing ratings, model-based
olutions typically show more accurate results [19] . The main char-
cteristic of model-based CF is the use of machine learning tech-
iques to learn models that are able to predict the missing rat-
ngs. In the last few years, different models have been proposed for
odel-based CF, such as Association rule-based [20,21] , Bayesian
M.G. Campana, F. Delmastro / Online Social Networks and Media 3–4 (2017) 75–97 79
n
[
M
s
t
m
u
o
M
c
p
f
w
i
r
b
T
i
l
g
f
R
w
u
a
b
m
U
w
r
t
i
n
s
a
i
a
o
r
r
q
m
i
l
i
t
(
s
e
a
M
i
a
3
e
s
t
c
a
s
p
t
r
c
f
i
r
b
r
d
s
d
u
c
d
t
R
u
a
a
t
a
p
e
c
H
d
3
n
l
f
i
h
j
i
i
r
i
etworks [22,23] , Support Vector Machines [24] , Neural networks
25] and, more recently, Deep learning methods [26] . However, the
atrix Factorization (MF) models [27] are considered to be the
tate-of-the-art in recommendation systems due to their advan-
ages with respect to scalability and accuracy [19] . Generally, MF
odels exploit the typically high correlation between rows (e.g.,
sers) and columns (e.g., items) of an incomplete ratings matrix in
rder to learn low-rank representations of both users and items.
oreover, these low-rank matrices (also referred as latent factors )
an be used to approximate the full rating matrix and, then, to
redict missing scores between users and items. Specifically, latent
actors derived from the correlation patterns in the ratings matrix,
hich can be used to describe with more details both users and
tems profiles. Consider the toy-example depicted in Fig. 2 . The
atings matrix contains the preferences of 3 users for 3 different
ooks. The latent factors may represent the genres of the books.
herefore, the users’ latent factors describe how much each user
s interested in a specific genre and, in the same way, the items’
atent factors represent how much a book belongs to a specific
enre.
Formally, the matrix factorization method can be expressed as
ollows:
≈ U · V
T , (11)
here R is the full rating matrix, and U and V are two matrices of
sers and items latent factors, respectively. The main goal of a MF
lgorithm is to learn the optimal latent factors ( U
∗ and V
∗) which
etter approximate R . The most common formulation of MF is for-
alized as follows [5] :
∗, V
∗ = argmin
U,V
{
1
2
M ∑
i =1
N ∑
j=1
I i j (r i j − U i V
T j )
2 +
λU
2
‖
U ‖
2 F
λV
2
‖
V ‖
2 F
}
,
(12)
here U i is the latent factors of user i and, in the same way, V i rep-
esents the latent factors of item i. I ij denotes an indicator function
hat is equal to 1 if rating r ij is not a missing value of the start-
ng rating matrix. Finally, ‖ ·‖ 2 F represents the (squared) Frobenius
orm of a matrix, and λU and λV are regularization parameters to
et in order to prevent the model’s overfitting. U and V can be seen
s unknown variables, which need to be learned in order to min-
mize the objective function presented in Eq. 12 . This is typically
chieved with the well-known Gradient Descent algorithm or one
f its variants (e.g., Stochastic Gradient Descent – SGD).
Compared to memory-based approaches, the predictions’ accu-
acy is not the only advantage of model-based CF. In fact, space
equirements for model-based CF are often lower than those re-
uired by memory-based algorithms. This is due to the fact that
emory-based approaches should load into memory all the rat-
ngs in order to perform the recommendations, while the model
earned by model-based CF is typically much smaller than the orig-
nal ratings matrix. However, learning a model may require lots of
raining data and time. In addition, if the system is highly dynamic
e.g., new users or items are frequently added), it could be neces-
ary to train a new model several times since the current one can
asily become obsolete, thus affecting its accuracy. This could be
n issue for some specific application domain. For example, in a
obile Social Network, each node (e.g., mobile device) has a lim-
ted knowledge of the items (and users) available in the network,
nd it could be difficult to train an accurate model.
.2. Content-based Recommender Systems
CF approaches can be considered inherently “social” since they
xploit the correlations in the ratings patterns among users of the
ame system. However, CF methods typically ignore the item’s at-
ributes to compute the predictions in their recommendation pro-
ess, while they can be useful to improve the recommendations’
ccuracy.
Content-based Recommender Systems [28,29] are specifically de-
igned to recommend items similar to the ones that the user has
referred in the past. In CF the similarity between two items (or
wo users) is calculated in terms of correlation or similarity among
atings provided by other users. Instead, content-based approaches
onsider only the ratings provided by the target user and the
eatures of the rated items [19] . Content-based RS are typically
mplemented using traditional classification and clustering algo-
ithms, such as Support Vector Machines (SVM) or Nearest Neigh-
ors methods (NN) [29] . In SVM the features of the items and the
atings made by the target user are combined together to form a
ataset of < features, rating > instances. Based on this dataset, a
pecific classifier (or regressor) for the target user is trained in or-
er to predict the ratings of new items, never seen before by the
ser. In NN, in order to classify a new, unrated item, the algorithm
ompares it with all the stored items, derived from the training
ata, by using a similarity function (typically the Euclidean dis-
ance or the Cosine similarity).
According to Aggarwal [19] and Tang et al. [16] , content-based
S have the following drawbacks:
1. Items’ features . The accuracy of RS relies on the set of features
that describe the items. The identification of the most relevant
features is not trivial, and highly depends on the specific appli-
cation.
2. Over-specialization . Since the content-based methods rely just
on the characteristics of the items already rated by the target
user, it typically suffers from low diversity and novelty of the
recommended items.
Thus, a content-based RS typically tends to provide obvious rec-
ommendations, which could bother the user in the long term.
3. Training set size . In order to allow a content-based RS to learn
the user’s preferences, the user has to rate a sufficient number
of items [16] . Otherwise, the RS does not has enough informa-
tion to learn an accurate model and it fails to recommend items
for a user with few or no ratings.
Despite the aforementioned drawbacks, content-based RS are
seful to alleviate some critical problems of CF. For example, when
new item is added to the ratings matrix, CF methods are not
ble to perform recommendations about this item because the sys-
em has not yet collected sufficient ratings about it. In the liter-
ture, this problem is called cold-start , and a content-based ap-
roach can well complement a CF system because of its ability to
xploit the characteristics of the items in the recommendation pro-
ess [19,30] . Mixing two or more approaches is often referred as
ybrid RS [31] in which the goal is to combine the strengths of
ifferent methods to create a more robust RS.
.3. Graph-based
Social Networks can be naturally modeled as graphs, in which
odes represent users and items, and edges model the different re-
ationships among user–user or user–item pairs (e.g., friendships,
ollow, likes, share). These relationships can also be used in RS to
dentify similar users and/or items. In fact, several graph-based RS
ave been recently proposed in the literature [32] . Their first ob-
ective is to perform a ranking of the nodes to identify the most
nteresting items for a user or a neighborhood (e.g., the most sim-
lar users/items) to be used in a traditional CF method.
PageRank [33] is clearly the most famous method to produce a
anking of the nodes in a graph. It is based on the idea that a node
s important if it is linked to other important nodes. This simple
80 M.G. Campana, F. Delmastro / Online Social Networks and Media 3–4 (2017) 75–97
b
s
t
t
c
e
d
w
a
t
o
h
a
P
p
o
p
p
r
o
w
e
a
g
b
l
t
i
d
f
u
i
H
λ
h
b
3
e
t
n
t
C
c
o
s
t
t
g
a
a
s
t
d
a
w
d
and general idea can be applied to many different situations, in-
cluding ranking of websites in a search engine, or finding relevant
items (or people) in a RS. PageRank exploits a random-walk model.
Specifically, it performs an exploration of the given graph by vis-
iting nodes following randomly selected links among them. The
long-term frequency of visits to a node represents its final rank-
ing value, which is clearly influenced by the number of incoming
links of the node. It is also referred to as the steady-state proba-
bility . However, directed graphs can have some nodes (or group of
them) without out-going links; this part of the network would act
as a trap for the random-walk process. In order to overcome this
problem, PageRank uses the so called restart mechanism. Accord-
ing to this strategy, at each transition, the random walk may either
jump to an arbitrary node of the network with a probability α, or
follow one of the out-going links connected to the current node
with probability (1 − α) . Formally, the steady-state probability at a
node i is defined as follows:
π(i ) =
α
n
+ (1 − α) ·∑
j∈ In (i )
π( j) · p ji , (13)
where n is the total number of network’s nodes, In ( i ) is the set of
nodes that have out-going links directed to i , and p ji is the transi-
tion probability from node j to node i .
Although PageRank is very effective in finding popular nodes
in the graph, it does not take into account the user’s preferences,
which means that it is not able to provide personalized recom-
mendations. To this aim, Haveliwala [34] proposes personalized
PageRank , a biased version of PageRank, specifically designed to
find popular nodes which are also similar to specific nodes in the
graph. In this case, the key idea is to use the restart mechanism to
bias the random walk process towards specific nodes that repre-
sent the user’s interests. If we consider a specific OSN represented
as a graph, nodes represent the items available in a social media,
and the links represent the relations between users and items. The
restart nodes represent the items consumed by the target user in
the past. Then, exploiting the biased random-walk, personalized
PageRank provides a way to calculate similarity scores of nodes
based on their structural similarity to the restart nodes. Finally,
the items with higher similarity values can be recommended to
the target user.
Even if personalized PageRank, and other similar random walk-
based algorithms (e.g., FolkRank [35] , SocialRank [36] , and Item-
Rank [37] ), can be directly employed to build Social Recommender
Systems (SRS) [36,38] , they are mostly used to select user/item’s
neighbors in memory-based CF solutions [39,40] . In fact, rather
than using standard similarity measures (e.g., Pearson Correlation
Coefficient on user’s ratings), it is possible to use such structural
measures to identify the most relevant users (or items) for the rec-
ommendation purpose. Since they only leverage on the structural
transitivity of the network’s edges, instead of the users’ ratings,
they typically result very effective to alleviate the problems related
to the sparsity of the ratings matrix. However, random walk-based
approaches typically require multiple iterations before the transi-
tion probabilities converge to a steady state.
On the contrary, other solutions require only few steps to
generate effective recommendations [41] . Diffusion-based methods
[32] represent the most used approach for recommendation in
graph-based RS. They have been initially proposed by Zhou et al.
[41,42] with their algorithms ProbS (Probability Spreading) [42] and
HeatS (Heat Spreading) [41] . These algorithms exploit a bipartite
user–item graph where users are directly connected to the items
collected in the past.
In ProbS, a resource is initially assigned to each item connected
to the target user, initialized to a unitary value. This resource is
then spread uniformly from the collected items to the users con-
nected with them ( first diffusion step ) and then, in a second step
ack, to the items connected with those users ( second diffusion
tep ). The final value of this resource, associated with each item, is
hen interpreted as the recommendation score of that item for the
arget user. The higher the score obtained by an item, the greater
ould be the interest in it by the target user.
HeatS [41] is similar to ProbS, but it is based on opposite rules:
ach time the resource (or a portion of it) is redistributed, it is
ivided by the number of edges connected to the node towards
hich it is heading to. Fig. 3 depicts the diffusion steps of the two
lgorithms, highlighting the differences in the two recommenda-
ions.
Unfortunately, both of them are actually biased by the presence
f extremely popular or non-popular items (i.e., nodes with very
igh or low degree, respectively), not taking into account the char-
cteristics of the users’ interests. To overcome these limitations, a
robS+HeatS hybrid approach, simply called Hybrid , has been pro-
osed in [41] . Hybrid calculates a linear combination of the results
f ProbS and HeatS with a parameter λ governing the relative im-
ortance of one of the two original algorithms. By tuning λ appro-
riately, Hybrid is able to obtain simultaneous gains in both accu-
acy and diversity of recommendations with respect to the previ-
us solutions.
The diffusion-based approach has been generalized in multiple
ays. Some authors proposed to compensate the HeatS’s prefer-
nce for unpopular items modifying its spreading process [43] and
llocating higher weight to links connecting nodes with high de-
ree [44] . On the other hand, in order to compensate the ProbS’s
ias towards very popular items, Lü and Liu [45] proposed to high-
ight less popular items in the second diffusion step by making
heir score inversely proportional to the degree of the nodes.
Furthermore, Zhou et al. [46] investigated the effect of bias-
ng the initial resource allocation in ProbS. Results indicate that
ecreasing the initial resource allocated on popular objects can
urther improve the algorithmic accuracy. Similarly, Liu and Zhou
sed heterogeneous initial resources in Hybrid, further improv-
ng both the accuracy and diversity of the recommendations [47] .
owever, in addition to the ProbS–HeatS hybridization parameter
, they introduce a second parameter to tune, which governs the
eterogeneity of the initial resources.
For a complete comparison and evaluation of the diffusion-
ased methods, we refer the reader to [32] .
.4. Context-Aware Recommender Systems
RS we have presented focus on recommending the most rel-
vant items to individual users taking into account just two en-
ities: users and items they liked in the past. However, they do
ot consider additional information that can be useful to improve
he recommendations’ quality and personalization [5] . To this aim,
ontext-Aware Recommender Systems ( CARS ) [48] have been re-
ently proposed in the literature. CARS methods perform their rec-
mmendations considering also the information characterizing the
pecific situation in which the target user is involved. This addi-
ional information is referred to as the context . The notion of con-
ext has been studied in multiple disciplines and several hetero-
eneous definitions exist [48] . In CARS, context is defined as the
dditional information relevant to improve recommendations, such
s the time of the day, the locations of the target user, and her
ocial relationships.
As we described in Section 2 , the objective of a traditional RS is
o approximate the rating’s function f R : U × I → R in order to pre-
ict the possible interest of a user for a specific item. CARS gener-
lize this approach by using a multidimensional approach [49] in
hich the rating function can be seen as a mapping from a n -
imensional tensor (i.e., a multidimensional matrix) to the set of
M.G. Campana, F. Delmastro / Online Social Networks and Media 3–4 (2017) 75–97 81
Fig. 3. Execution of ProbS and HeatS on a bipartite user–item graph.
Fig. 4. Example of a multidimensional cube for the User × Item × Location recom-
mendation space.
r
w
(
v
a
d
s
t
h
a
fi
I
C
d
d
r
r
b
l
a
e
T
t
i
c
L
c
s
m
t
a
(
d
d
p
u
t
u
b
t
t
a
p
u
u
E
t
r
c
v
t
c
C
t
T
c
atings R :
f R : D 1 × · · · × D n → R, (14)
here two dimensions, typically, are the sets of users and items
i.e., U and I ), and the other dimensions represent the context
ariables. Grafically, CARS’s rating function can be represented as
multidimensional cube. Fig. 4 shows an example with a 3-
imensional cube that stores the ratings for the recommendation
pace User × Item × Location , where f R (u 1 , i 4 , home ) = 5 means that
he user u 1 rated with a score of 5 the item i 4 when she was at
ome.
The paradigms used to incorporate context information in RS
re divided in three main categories [50] : (i) Contextual pre-
ltering , (ii) Contextual post-filtering , and (iii) Contextual modeling .
n the following we describe in details each approach.
ontextual pre-filtering. In contextual pre-filtering the idea is to re-
uce the multidimensional cube to a 2-dimensional matrix in or-
er to apply traditional algorithms. As depicted in Fig. 5 , the cur-
ent context c of the target user is used as a filter to “slice” the
ating cube and extract the relevant user–item matrix, which can
e used with traditional CF algorithms.
However, using the exact context c sometimes can be a prob-
em. If there are just few ratings associated with c , this can lead to
low accuracy in the recommendation because the system has not
nough data points about the past preferences of the target user.
herefore, the notion of generalized pre-filtering [49] has been in-
roduced in the literature to cope with this problem. Instead of us-
ng the specific context c , this approach suggests to use a broader
ontext c ′ , which includes c and other similar context variables.
et us consider the context of watching a movie on Saturday (i.e.,
= “Saturday”). In order to base the recommendation on a wider
et of ratings, the generalized pre-filtering approach may use the
ore general context c ′ = “Weekend" , which includes the two con-
ext variables c = “Saturday” and c 1 = “Sunday”. Then, the system
ggregates the selected ratings using some aggregation function
e.g., the average) to reduce the recommendation space to the 2-
imensional user–item matrix and, finally, perform the recommen-
ation.
It is clear that the context generalization becomes a crucial as-
ect of the recommendation process. One possible strategy is to
se a process similar to the cross-validation: in a learning phase,
he predictive performance of the system is empirically evaluated
sing different contextual filters. Then, the filter that allows the
est performance is automatically selected [50] . In addition, mul-
iple context generalization can be used for the same specific con-
ext. For this reason, in the literature several works propose to use
combination of different filters and prove that this solution can
rovide a significant performance improvement compared to the
se of a single filter [50,51] . In this case, multiple pre-filters are
sed to generate several different 2-dimensional ratings matrices.
ventually, those user–item matrices are merged in one single ma-
rix in order to perform the recommendation process. Codina et al.
ecently proposed in [52] the use of semantic similarities between
ontextual situations in order to select the most relevant context
ariables for a specific situation. They showed that their seman-
ic pre-filtering model is able to obtain better results than others
ontext-aware approaches.
ontextual post-filtering. In this case the filtering step is applied to
he output obtained after applying traditional RS to the data set.
herefore, the recommendation process of a post-filtering method
an be summarized in two steps:
82 M.G. Campana, F. Delmastro / Online Social Networks and Media 3–4 (2017) 75–97
Fig. 5. Process of a pre-filtering method to generate contextual recommendations.
Fig. 6. Process of a post-filtering method to generate contextual recommendations.
t
t
p
c
a
a
c
m
f
(
c
r
w
m
t
h
p
a
(
f
(
c
t
w
i
b
s
a
b
s
F
m
1. In a first stage, the context information are completely ignored,
and a list of recommended items are generated using tradi-
tional RS on an aggregated 2-dimensional user–item matrix.
2. Given a specific context c , recommendations not relevant for c
are filtered out, and the ranking of recommendations in the list
is adjusted according to c .
Fig. 6 depicts the recommendation process of the post-filtering
approach. Here, aggr ( U, I ) refers to the aggregation function (e.g.,
the average) used to create the 2-dimensional user–item matrix,
and filter ( c ) represents the filtering (or adjustment) step of the rec-
ommendation’s list, based on the current context c of the target
user.
The selection of the best contextual filter represents a crucial
point in the overall recommendation process. All the techniques
related to the context generalization, previously discussed, can be
applied also in post-filtering methods. The main difference be-
tween the two approaches lies in the data set on which the con-
textual filter is applied: the ratings contained in the multidimen-
sional tensor (i.e., the input) for the pre-filtering approach, and the
recommendation list (i.e., the output) generated by a traditional RS
for post-filtering methods.
Both approaches share the same advantages: their implemen-
tation is relatively simple, and both of them allow the use of the
traditional recommendation techniques described in the previous
sections. In addition, comparing their performance, none of them
seems to clearly outperform the other. Panniello et al. [53] evalu-
ated different solutions based on both approaches by using several
real world datasets. The authors did not find any solution that
dominates over the others for all the datasets, indicating that
the choice of the best approach really depends on the specific
application.
Contextual modeling. Contextual modeling approaches consider
context information directly in the recommendation function in
order to predict a user rating for a specific item [50] . Differently
from pre- and post-filtering techniques, contextual modeling meth-
ods use predictive models or heuristics to create multidimensional
recommendation functions.
The simplest solution adapts neighborhood-based CF methods
o perform context sensitive recommendations [49] . In this case,
he traditional user–user (or item–item) similarity metric is re-
laced by an n -dimensional distance metric, which includes the
ontextual information. For instance, let’s consider two points in
3-dimensional user–item–context cube, respectively A = (u, i, c)
nd B = (u ′ , i ′ , c ′ ) . The distance between them, Dist ( A, B ), can be
alculated as the sum of the weighted distance between each di-
ension or, alternatively, using the Euclidean metric [49] . There-
ore, by extending the traditional neighborhood rating prediction
i.e., Eq. (7) ), the specific rating for user u , item i in the context c ,
an be expressed as follows:
u,i,c = k ∑
(u ′ ,i ′ ,c ′ ) � =(u,i,c)
1
Dist((u
′ , i ′ , c ′ ) , (u, i, c)) · r u ′ ,i ′ ,c ′ , (15)
here k is a normalization factor, and the inverse of the multidi-
ensional distance between points has been used for weighting
he rating r u ′ ,i ′ ,c ′ (i.e., the more the two points are close, and the
igher is the weight).
More complex approaches extend model-based solutions. In
articular, tensor factorization methods [54,55] can be considered
s a generalization of matrix factorization for latent factor models
see Section 3.1.2 ). As depicted in Fig. 7 , a 3-dimensional tensor
actorization aims at learning three latent factors matrices ˆ U , ˆ I , C
i.e., respectively, the user, item, and context latent factors) and a
ore tensor ˆ T to build a prediction model. However, the compu-
ational complexity of this kind of models increases exponentially
ith the number of the tensor’s dimensions [56] . To solve this
ssue, some simplified forms of these factorization methods have
een recently proposed. For instance, the Pairwise Interaction Ten-
or Factorization ( PITF ) [56] and Factorization Machines [55,57] are
ble to predict missing ratings exploiting the pairwise interactions
etween the different dimensions of the tensor, maintaining at the
ame time a low computational complexity.
To conclude the description of main RS methods, we present in
ig. 8 a summary of their classification and the references of the
ain representative solutions for each method.
M.G. Campana, F. Delmastro / Online Social Networks and Media 3–4 (2017) 75–97 83
Fig. 7. Tensor factorization.
Fig. 8. Classification of Recommender Systems methods.
4
e
i
i
s
p
o
i
i
c
o
t
i
(
o
i
w
t
s
o
t
p
. Recommender systems for Online Social Networks
With the advent of OSN, RS have been further enhanced by
xploiting additional information characterizing both users and
tems, becoming thus SRS [16,58] . Specifically, OSN can provide
nformation able to identify virtual and physical social relation-
hips among users, common interests and habits, in addition to
ersonal preferences. RS can exploit this information depending
n the application domain in which they are implemented. Specif-
cally, SRS widened their recommendation targets, from generic
tem recommendations up to people/friend recommendations, lo-
ations, tags and others. In this section, we provide a description
f SRS grouped by the type of recommendations they provide, and
he context information used to optimize their process, as shown
n Fig. 9 . Specifically, we identify three types of recommendations:
i) social-aware recommendations , in case SRS recommend people
r friends, and include the notion of trust and social relationships
n generic items recommendation; (ii) tag-based recommendations ,
hen SRS exploit a specific characterization of items based on tags
o recommend other items, tags and/or people, e.g., if they share
imilar interests derived from tags; (iii) location-aware rec-
mmendations , when SRS exploit location-related information
o recommend items, point-of-interest (POIs), trajectories, and
eople.
84 M.G. Campana, F. Delmastro / Online Social Networks and Media 3–4 (2017) 75–97
Fig. 9. Classification of Social Recommender Systems.
a b
Fig. 10. Graphical representation of (a) a basic friendship network, and (b) a trust-network with inferred trust links (dashed lines).
I
b
o
g
m
p
s
4
o
t
p
t
n
t
e
l
t
fi
t
(
F
4.1. Social-aware recommendations
OSN provide several information that characterizes the social
context of a user and her generated items. OSN analysis mainly
refers to the topological structure of the social network, in terms of
(virtual) social relationships among users [59,60] . In addition, OSN
can reflect the presence of the homophily principle [61] also in the
cyber-world, as the tendency of individuals to associate and con-
nect with similar others, as demonstrated by recent studies [60] .
The social structure of a OSN can be represented as a graph
G = (U, S) , where U is the set of users and S is the set of social
links among them, which can be: (i) undirected, if they model
friendship relations, such as in Facebook, or (ii) directed, if they
model trust or follow relations, such as in Epinions 3 or Twitter.
Social links in OSN represent the explicit declaration of a user to
be virtually in contact with others, but they do not reveal anything
about the effective nature of the relationships or its strength and
value for the user. However, analyzing OSN contents and users’ vir-
tual interactions (e.g., explicit users’ feedbacks, ratings [62] , etc.)
we can derive additional information that can further character-
ize those links, like for example the notion of trust among users,
which can be represented through links’ weights ( Fig. 10 (b)).
SRS exploit this structure to provide two types of social-aware
recommendations: (i) friends or other people, and (ii) OSN items.
3 http://www.epinions.com
[
b
b
n the first case, SRS suggest to the users new social relationships
ased on the prediction of common interests, while in the sec-
nd case they exploit social/trust relationships in order to sug-
est interesting items to the users. In the following we review the
ain solutions proposed in the literature, highlighting different ap-
roaches, and how they take into account different aspects of the
ocial context.
.1.1. Friends and people recommendations
This recommendation task differs from the generic items rec-
mmendation since it must take into account specific aspects of
he social network of each user, like trust, reputation, privacy and
ersonal attraction [58,63] . From a technical point of view, this
ask is associated with the problem of link prediction in the social
etwork, which tries to infer new possible relationships or interac-
ions between pairs of entities, based on their properties and the
xisting links [64] .
In the literature, several techniques have been proposed for
ink prediction [65] . The most commonly used can be divided into
wo main categories: unsupervised and supervised methods. The
rst category includes all those techniques that exploit the struc-
ure of the network to calculate similarity values among the users
nodes of the graph), and then recommend the most similar ones.
or instance, Jaccard coefficient [65] or the Adamic–Adar measure
65] can be used to calculate the similarity between two nodes
ased on their common neighbors. Alternatively, random walk-
ased approaches such as Katz measure [65] , or PageRank [33] and
M.G. Campana, F. Delmastro / Online Social Networks and Media 3–4 (2017) 75–97 85
S
m
h
n
b
r
c
a
fi
w
e
i
b
o
c
a
n
p
e
(
d
i
c
m
d
s
o
t
i
a
(
c
a
c
r
a
h
a
b
n
m
o
n
s
t
a
d
b
t
m
t
o
t
p
d
o
t
s
a
t
a
r
p
o
p
r
s
i
b
f
(
i
s
u
W
t
l
i
t
(
i
h
d
t
s
t
t
p
s
e
p
l
f
4
m
h
d
i
v
p
n
o
f
s
r
s
(
a
C
u
t
o
i
n
f
g
n
s
i
g
t
o
v
imRank [66] algorithms, can represent a different way to recom-
end new links based on the connectivity between pairs of nodes.
One of the first unsupervised RS for people recommendation
as been proposed by Guy et al. for an IBM enterprise social
etwork called SONAR [67–69] . It defines a “relationship score”
ased on a combination of various social context information de-
ived from the enterprise intranet applications (e.g., organizational
harts, papers database, patents database, a social networking site,
friending system and others). For each application, they de-
ne a user–user graph in which the links’ weight is computed
ith a specific metric depending on the application (e.g., manager-
mployee relationship, paper/patent co-authorship, users belong-
ng to the same group in the friend network). Then, a linear com-
ination of the weights of each graph is computed for each pair
f users, and it is used as general relationship score. The authors
ompared SONAR relationship score with that provided by other
pproaches: (i) Friend-of-friends (FOF) metric, applied to the inter-
al friending system of the company; (ii) a Content-Matching ap-
roach, defining the users’ similarity based on a set of keywords
xtracted only from the contents generated on the social network
i.e., posts, picture descriptions and tags); (iii) Content-plus-Link , by
oubling the similarity value derived from the Content-Matching
n case the pair of users are also explicitly connected in the so-
ial network application. They evaluated SONAR with real experi-
ents involving up to 30 0 0 users in different testing scenarios and
emonstrated the effectiveness of the proposed solution with re-
pect to the other approaches.
Other solutions have been proposed by exploiting the concept
f ego-network , the representation of a social network centered on
he individual (i.e., ego) and her social links (i.e., alters) divided
n a series of “circles”. The inner-most circle includes alters with
very strong relationship with the ego. Each subsequent circle
in hierarchy) includes all the relationships of the previous cir-
les along with an additional set of social links characterized by
weaker level of intimacy. The last set, included in the outermost
ircle only, contains simple acquaintances, with a relatively weak
elationship with the ego. This representation has been defined in
ntropology for classical (offline) social networks [70,71] , and it
as been recently demonstrated that the same representation can
lso be adopted in OSN [72,73] .
Epasto et al. [74] proposed a RS for friends recommendation
ased on the features derived from the co-occurrences of two
odes in different ego-networks. In contrast to random walk-based
easures, these features can be computed much more efficiently
n very large graphs, by just analyzing selected neighbors of each
ode. The authors evaluated the system on an anonymized snap-
hot of Google+ social network, and compared the proposed solu-
ion with two other unsupervised methods (i.e., common friends
nd Adamic–Adar) in terms of precision and recall. A 2-week test,
uring which the authors observed the social network evolution
ased on the different recommendations, demonstrated the effec-
iveness of their solution.
In general, the effectiveness of the proposed unsupervised
ethods highly depends on the social network structure used for
heir evaluation, and they are not easy to compare [75] . On the
ther hand, supervised methods try to cope with this issue, by
reating the link prediction in a social network as a classification
roblem and by defining the characteristics of the network in a
ata-driven way, i.e., by creating a vector of features for each pair
f nodes, in which the set of features can include different heuris-
ics and metrics. To this aim, for each pair of nodes, a multidimen-
ional vector is extracted and associated with a positive or neg-
tive label, based on the presence or absence of a link between
he nodes. Finally, a classifier is trained using the set of vectors
s training data. In this way, the classifier automatically learns the
elevance of the features, and the learned model can be used to
redict the creation of new links (i.e., the label) between any pair
f nodes.
The following solutions are representative of this class of link
rediction approaches. Fire et al. [76] proposed a set of features
elated to different characteristics of the graph (e.g., nodes, edges,
ubgraphs, and multi-hop paths) in order to learn a classifier that
dentifies the missing links. They used different datasets (e.g., Face-
ook, Flickr, YouTube, Academia, etc.) and by using a common
ramework they developed different classifiers for each dataset
e.g., decision tree C4.5, kNN, Naive Bayes, SVM, neural networks)
n order to evaluate both single solutions and the combinations of
ome of them. Finally, they compared the prediction accuracy by
sing all the features or some limited sets.
Backstrom and Leskovec [77] proposed Supervised Random
alks, a supervised algorithm that combines the network struc-
ure with attributes about both nodes and edges in a unified
ink prediction algorithm. Given a source node s, and train-
ng examples about which nodes s will connect to in the fu-
ure, the algorithm learns a function that assigns a strength
i.e., random walk transition probability) to each edge, provid-
ng higher scores to the possible new links of s . This solution
as been compared to both unsupervised methods (e.g., ran-
om walk with restart, Adamic–Adar, Common Friends) and other
wo supervised approaches (i.e., decision-tree e logistic regres-
ion) on heterogeneous datasets (i.e., arXiv and Facebook). All
he supervised methods outperformed the unsupervised ones in
erms of precision and the ranking metric AUC [108] , and Su-
ervised Random Walk performed better than the others of the
ame class. We can thus observe that supervised methods gen-
rally performs better than unsupervised ones in friends and
eople recommendations, thanks to their ability to automatically
earn the hidden relationships between social links and their
eatures.
.1.2. Items recommendations
Social context positively impacts also on generic items’ recom-
endation (such as books, movies, or news), and several works
ave been presented in the literature exploiting most of the stan-
ard recommendation techniques. The first category we consider
s related to the neighborhood CF approach. These solutions use
arious strategies to select relevant users for the recommendations
urpose. For instance, authors in [78,79] considered only the con-
ected friends of the target user and their items as the possible
bjects to recommend. Instead, in [79] an evaluation between dif-
erent strategies has been proposed. Specifically, the authors con-
idered (i) a Friends Average approach, which simply averages the
atings of the direct friends, (ii) a Weighted Friends, in which co-
ine similarity among friends is used to define links’ weight, and
iii) a traditional CF method without social context. All the social-
ware recommendation tasks achieve better results than traditional
F in terms of prediction accuracy.
Other memory-based solutions exploit trust relations among
sers to obtain the set of relevant neighbors [80–82] . Generally,
hese solutions exploit different graphs’ exploration techniques in
rder to propagate trust within the user-to-user network. In fact,
n this case, the network is represents as a direct graph in which
odes represent the users and edges connect two users if they are
riends or at least one of them declares a trust towards the other
reater than 0. As a first step, the trust values among directly con-
ected users are typically calculated leveraging on a user-to-user
imilarity measure (e.g., the cosine similarity calculated on the rat-
ngs of the two users) [16] . Then, these values of trust are propa-
ated through the graph’s edges in order to calculate the transi-
ivity trust among nodes that are not directly connected to each
ther. For instance, TidalTrust [80] aims at estimating the trust
alues of a node (i.e., source) towards another one not directly
86 M.G. Campana, F. Delmastro / Online Social Networks and Media 3–4 (2017) 75–97
p
b
t
F
p
r
p
f
[
b
m
s
r
4
w
t
i
s
t
o
O
t
h
k
c
t
i
a
F
a
s
f
S
l
t
p
m
o
t
h
w
s
t
s
n
t
t
a
h
s
A
s
s
Y
e
connected (i.e., sink). In a first phase, it finds a path from the
source to the sink while rating the nodes on the path, and then
it aggregates the trust value backwards towards the source. As a
result, TidalTrust assumes that: (i) shorter propagation paths pro-
duce more accurate trust estimates, and (ii) paths with higher trust
values create better results in selecting neighbor nodes to be used
for items recommendations. Therefore, TidalTrust selects as rele-
vant neighbors only those nodes that are on the shortest path be-
tween the source user and the sink. MoleTrust [81] is another ex-
ample, which uses a random walk to propagate trust values on the
graph, and it defines a threshold on the length of the path between
a source and a sink. Therefore, it limits the number of possible
paths between a pair of users.
In contrast with memory-based approaches, model-based Rec-
ommender Systems learn models that predict the user’s interest
for new items. As it happens for traditional RS, the matrix factor-
ization (MF) represents the most used model-based technique in
SRS. The common idea of Social-aware MF solutions is that social
connections mainly influence the users’ preferences (i.e., the rat-
ings). The preferences of two friends should be much more similar
than those of two strangers, and for this reason these solutions
typically associate a weight to the social relations, which indicates
the strength of the social tie among the related users and it is usu-
ally calculated based on the ratings similarity.
Each of the MF solutions proposed in the literature differs from
the others by the method used to include social and/or trust infor-
mation into the factorization process. For example, SoRec [83] , in-
stead of factorizing only the user–item matrix, as in traditional MF
methods, it extends the factorization objective function performing
a co-factorization both of the user–item matrix and the user–user
social matrix. Instead, other solutions (e.g., SocialMF [84] and Social
Regularization [85] ) focus on the regularization term of the factor-
ization objective function (see Section 3.1.2 ), introducing new con-
straints related to the social factors. Specifically SocialMF [84] , by
assuming that the behavior of a user is affected by the behavior
of her direct neighbors, introduces a new regularization term that
forces the preference of a user to be closer to the average prefer-
ence of the users in her social neighborhood. On the other hand,
Social Regularization [85] considers the fact that the target user’s
neighbors may have different tastes. For this reason it uses a pair-
wise regularization of the users’ latent factors where the prefer-
ences’ similarity of two connected users depends on their ratings’
similarity. In order to evaluate Social Regularization, authors com-
pared the performance of their solution with other three MF ap-
proaches: two traditional user–item MF (i.e., non-negative matrix
factorization [86] , probabilistic matrix factorization [87] ), and So-
cial Trust Ensemble (STE) [88] , a social-aware MF solution where
the missing ratings are predicted as a linear combination of rat-
ings from the target user and her friends. By using the datasets
extracted from two real-world OSNs (i.e., Douban
4 and Epinions),
Social Regularization demonstrated to outperform the other ap-
proaches in terms of both MAE and RMSE metrics. In addition,
in the same work [85] , authors propose the use of two differ-
ent similarity metrics for the pairwise regularization: the Pearson
Correlation Coefficient and Vector Space Similarity [89] . However,
the evaluation results show no significant differences between the
two measures, but they both perform better than random similar-
ity values, showing the importance of similarity function enhanced
with social context information.
Recent works (i.e., [90,91] ) use different regularization terms.
In [90] , Sun et al. propose RSboSN (Recommender System based
on Social Networks), which leverages on a clustering algorithm to
identify groups of friends. RSboSN outperforms SoRec in terms of
4 https://www.douban.com
t
c
t
recision and recall, proving the effectiveness of their clustering-
ased regularization approach. Moreover, Reafee et al. [91] propose
he EISR (Explicit and Implicit Social Relation Probabilistic Matrix
actorization) algorithm, aimed at including both explicit and im-
licit social relations as regularization parameters. Explicit social
elations are the direct friendships ties among users, while the im-
licit ones refer to not directly connected users and they are in-
erred using link prediction. Compared with Social Regularization
85] and two diffusion-based algorithms (i.e., HeatS [41] and Hy-
rid [41] , described in Section 3.3 ), EISR shows a clear improve-
ent in terms of prediction error (i.e., MAE and RMSE), demon-
trating the effectiveness of taking into account also the implicit
elationships in the social recommendations.
.2. Tag-based recommendations
Over the past few years the use of special user-defined key-
ords, called tags , to categorize or describe web and OSN con-
ents, gained a lot of popularity. This user-driven phenomenon
s known in the literature as folksonomoy [92] and it is a well-
tudied topic in both information retrieval and Recommender Sys-
ems fields [93–95] .
An important aspect of folksonomies is that, differently from
ntologies, no relationship between the terms is required a priori.
n the contrary, relationships are automatically built exploiting the
ags created by the users and explicitly assigned to contents (e.g.,
ashtags assigned to tweets). Since tags are generally defined as
eywords that can reflect a semantic meaning of the associated
ontent, they represent an important feature in the content charac-
erization. In addition, folksonomies dynamically adapt to changes
n the users’ vocabulary, further personalizing the user’s interests,
nd they are independent from the type of item they describe.
or instance, the features used to represent the characteristics of
movie are typically very different from the features used to de-
cribe music files or books. Instead, tags can be seen as generic
eatures that can be used to create multi-domain Recommender
ystems [96] .
However, folksonomies have some drawbacks due to the user
anguage. Synonyms, homonyms, polysemies, and different users’
agging behavior make the recommendation process difficult to
erform in some cases, and undermine the use of simple tag
atching (e.g., group items characterized by the same tag). To
vercome these limitations, a novel family of Recommender Sys-
ems, called Tag-based Recommender Systems (TBRSs) [94,95] ,
ave been proposed in the literature.
Formally, the folksonomy can be modeled in the following two
ays. First, it is possible to use three distinct sets of elements: the
et of users U = { u 1 , . . . , u n } , the set of items I = { i 1 , . . . , u m
} , and
he set of tags T = { t 1 , . . . , t k } . These elements can represents the
et of nodes in a tripartite graph, as shown in Fig. 11 (a). Each bi-
ary relation between them can be described using adjacency ma-
rices, A
UI , A
IT , and A
UT for user–item, item–tag and user–tag rela-
ions respectively. If the user u l has collected the item i s , we set
UI ls
= 1 , otherwise a UI ls
= 0 . Similarly, we set a IT sq = 1 if the item i s
as been tagged with t q, and a IT sq = 0 otherwise. Furthermore, we
et a UT lq
= 1 if u l owns items tagged with t q, and a UT lq
= 0 otherwise.
lternatively, the folksonomy can be modeled as a third-order ten-
or taking into account only complete ternaries Y = (u, i, t) , as
hown in Fig. 11 (b). Specifically, the tensor’s component is set
= 1 if the ternary exists in the folksonomy, and Y = 0 otherwise.
A variety of different approaches have been proposed in the lit-
rature for tag-based RS. Works presented in [97,98] investigated
he tag co-occurrence frequency to support the user tagging pro-
ess. Specifically, given a specific item (e.g., a photo in Flickr),
he system recommends to its users the most relevant tags to
M.G. Campana, F. Delmastro / Online Social Networks and Media 3–4 (2017) 75–97 87
a b
Fig. 11. Representation of a folksonomy with (a) a graph, and (b) a tensor.
a
f
c
q
u
t
t
i
n
t
(
o
c
o
b
p
g
f
f
r
s
i
o
t
a
l
p
e
t
h
r
t
n
e
p
v
[
b
w
b
(
u
s
i
R
S
b
c
u
t
d
c
t
n
I
p
m
p
l
m
u
i
t
n
t
f
t
f
c
t
w
n
o
i
i
a
t
a
o
d
[
t
t
r
p
t
i
o
i
t
h
m
m
p
p
4
w
o
F
5 https://foursquare.com/ 6 https://www.yelp.com
ssociate with it. The underlying idea is that, if the co-occurrence
requency of two tags is high, they could be closely related. On the
ontrary, if two tags are not related at all, their co-occurrence fre-
uency should be very low.
Several works extend CF towards tag-based RS. To this aim, the
ser–item–tag ternary is typically reduced to two-dimensional ma-
rices. Firan et al. [99] leverage on the user–tags projection matrix
o compute a ranked list of tags, then a list of recommended items
s extracted according to those tags. Instead, Guy et al. [100] select
eighbors users by using a similarity measure based on common
ags and items, while [101] exploits the co-occurrence of tags.
In order to overcome the issues related to the folksonomies
e.g., synonyms and polysemies), and to improve the performance
f CF, Shepitsen et al. [102] and Gemmell et al. [103] use hierarchi-
al tags clustering techniques, while [104] weights each tag based
n the relationships between users, items, and tags. Recently, a hy-
rid framework for both item and tag recommendations has been
roposed in [96] . Authors use text analysis techniques (e.g., bi-
rams) to solve the ambiguity problem of folksonomies and per-
orm tag recommendations. In addition, they exploit temporal in-
ormation and trust relationships among the users to provide item
ecommendations using a CF approach.
By modeling the folksonomy as a third-order tensor, some re-
earchers proposed Recommender Systems based on tensor factor-
zation techniques (see Section 3.4 ). Specifically, in [105] the rec-
mmendation is performed by the product of three low rank ma-
rices (i.e., the latent components of the users, items, and tags)
nd a low rank core tensor produced by the Higher-Order Singu-
ar Value Decomposition (HOSVD) [106] . Rendle et al. [107] pro-
osed a different learning approach for tensor factorization mod-
ls called Ranking with Tensor Factorization (RTF), which optimizes
he model parameters for the ranking statistic AUC [108] .
Another natural way to represent the folksonomies is to use
ypergraphs. According to this representation, nodes of the graph
epresent the folksonomy entities (i.e., users, items, and tags), and
he relationships between them are modeled with edges among
odes. Consequently, a number of researchers proposed differ-
nt graph-based solutions for tag-based RS. These approaches ex-
loit the structure of the folksonomy hypergraph in order to pro-
ide tags or items recommendations to the user. Hotho et al.
93] proposed FolkRank, a graph-based ranking algorithm inspired
y PageRank [109] . Here, the underlying idea is that an item tagged
ith “important” tags by “important” users becomes “important”
y itself. Essentially, FolkRank identifies the most “important” tags
i.e., the most visited tags during the random walk) for a specific
ser by performing a biased random walk with restart on the folk-
onomy graph.
Other solutions refer to the diffusion-based approach described
n Section 3.3 . Zhang et al. [110] first proposed a diffusion-based
ecommender System based on the tripartite user–item–tag graph.
pecifically, they apply the diffusion process proposed in [42] on
oth the user–item and item–tag bipartite networks and then, they
ombine the two results to provide the recommendations to the
ser. Their results show that the integration of tags can enhance
he accuracy of the recommendations with respect to previous
iffusion-based algorithms. However, due to their diffusion pro-
ess, the solution proposed in [110] and other similar works tend
o suggest the most popular items in the folksonomy network, pe-
alizing the personalization of the recommendations [41] .
To this aim, we recently proposed PLIERS (Popularity-based
tEm Recommender System), a tag-based RS designed both for bi-
artite [111] and tripartite graphs [112] . In order to provide even
ore personalized items to the users, PLIERS assumes that the
opularity of an item/tag (i.e., the degree of the node) can be re-
ated to its semantic. Therefore, a very popular item or tag can se-
antically relate to a more “generic” topic compared to a less pop-
lar item/tag that, instead, describes a more “specific” topic. For
nstance, any content related to the football club Millwall can be
agged with both tags “Millwall” and “Football”, but the opposite is
ot always true: all content concerning football will not always be
agged with “Millwall”. According to this assumption, we can there-
ore say that the tag “Football” refers to a more generic topic than
hat referred by the tag “Millwall”. Users interested in the Millwall
ootball club, but not connected to items tagged with “Football”, are
learly not interested in all the items tagged with the latter tag, as
hese could contain information about other football clubs. In this
ays, PLIERS solves the dilemma of the choice between popular or
on-popular items in the network in a more natural way than the
ther diffusion-based algorithms. PLIERS does not require any tun-
ng parameter, and it ensures that the popularity of recommended
tems is always comparable with the popularity of items already
dopted by the users. Compared with other diffusion-based solu-
ions, PLIERS shows comparable performance in terms of precision
nd recall, while providing better novelty in the recommendations.
A very recent and promising approach to build tag-based Rec-
mmender Systems is represented by the use of deep learning to
iscover latent features from the tag space. The system proposed in
113] uses a sparse autoencoder [114] to extract a set of dense la-
ent features from the user–tag matrix. Based on the extracted fea-
ures, users’ profiles are updated and the traditional CF is used to
ecommend items or tags. On the other hand, Xu et al. [115] pro-
oses a pure deep learning model to provide item recommendation
o the user. Here, two neural networks map the tag-based user and
tem profiles to an abstract deep feature space. Then, the relevance
f the specific item for the user is calculated based on the sim-
larity between the deep representation of the two profiles. Au-
hors compared their solution with three different approaches: a
ierarchical clustering model [102] , CF and the autoencoder-based
odel proposed in [113] . With respect to these solutions, experi-
ental results show that the proposed approach significantly im-
roves the effectiveness of the recommendations in terms of both
recision and recall.
.3. Location-aware recommendations
The recent advances in location-acquisition techniques, and the
ide spread use of GPS-enabled smartphones, led to the creation
f the so-called Location-Based Social Networks (LBSN), such as
oursquare 5 or Yelp
6 . These systems allow users to tag, rate, and
88 M.G. Campana, F. Delmastro / Online Social Networks and Media 3–4 (2017) 75–97
Fig. 12. Data in location-based social networks.
u
C
n
b
[
n
w
f
d
c
a
m
t
m
p
“
fi
[
f
s
n
c
c
(
a
P
c
w
p
c
i
D
i
m
c
w
t
c
a
t
i
o
s
f
l
s
e
i
m
u
t
i
f
o
e
p
m
e
t
o
t
m
e
describe the locations they visit, in order to aid the discovery of
unknown points of interests (POIs), like restaurants or shops, or to
enable new social relations among users [116] .
According to Bao et al. [117] , the location is one of the most im-
portant information to describe the user’s context. The addition of
spatial information in social networking systems is able to bridge,
at least partially, the gap between cyber and physical social rela-
tionships among users. Fig. 12 depicts the main information mod-
eled by a typical LBSN. In the upper layer, users and their social
relationships are represented. These social relationships can be ex-
plicitly defined by the users (e.g., friends or followers) or they can
be inferred from their activities on the OSN. The middle layer de-
picts the social media content generated by the users. The nature
of these content can be very different; for instance, a user can
share a check-in at a public place, a geo-tagged photos, a review
of a restaurant, or the workout tracked by a fitness application. As
depicted in the lower layer, each content should be associated with
one or more geographical locations.
Given the multi-dimensional nature of the LBSNs, SRS applied
to these specific OSN must be able to handle a multitude of het-
erogeneous data. They can be divided into four different categories
based on their recommendation task: (i) location, (ii) user, (iii) ac-
tivity, and (iv) social-media. The first category mainly focuses on
suggesting new locations to the user, based on her preferences.
The objective of the recommendations can be both single POIs or
a sequence of locations, such as the best route to take or the se-
quence of attractions to visit during a trip. To recommend a single
place, both content-based and CF techniques have been exploited.
In the former case, the solutions proposed in the literature, e.g.,
[118–120] , typically suggest new locations by matching the user’s
profile with the places’ meta-data (e.g., text description, categories,
or tags). These solutions inherit the main advantages and draw-
backs of the content-based approach: they are effective in mitigat-
ing the cold-start problem but, on the other hand, they may suffer
from the recommendation quality issue. In fact, taking into account
just the preferences of the target user and ignoring the opinions
coming from other users, the system may recommend a matching
place with poor quality from the social opinion standpoint. In or-
der to overcome this problem, a large number of works exploit the
users’ location histories and CF models. Leveraging on the other
users’ past actions (e.g., reviews or check-ins), these location-based
RS are able to improve the quality of the recommendations by ig-
noring poorly-reviewed locations that otherwise match the target
ser’s profile. As we argued in Section 3.1 , the crucial point in the
F’s recommendation process is represented by the selection of
eighbors users (or items). The early solutions define their neigh-
ors selection according to the first law of geography of Tobler
121] , which claims that “everything is related to everything else, but
ear things are more related than distant things”. For instance, the
ork in [122] considers only individuals who live near the location
rom which the target user has made the query, while in [123] can-
idate POIs are pruned by using a spatial range around the user’s
urrent location. In addition to the spatial distance between places
nd users, several works propose to use a different kind of infor-
ation in order to further improve the quality of the recommenda-
ions. For example, Shi et al. [124] proposes a category-regularized
atrix factorization model, taking into account the category of the
laces visited by the user in the past (e.g., “Italian restaurant” or
Irish pub”).
Sentiment analysis can also be used in order to build a more
ne-grained model of the user’s preferences [125,126] . Yang et al.
127] exploit the user’s comments left in the check-ins in a tensor
actorization framework to build a personalized location ranking
ystem, while the system proposed in [126] considers as neighbors
ot just the users with similar preferences, but also those with
omparable needs. Specifically, it clusters the places’ reviews ac-
ording to different aspects, such as the topics covered in the text
e.g., service, food, facilities), the intent and the nationality of the
uthor. In a recent work, Wang et al. [128] face the problem of
OIs recommendation considering the user interest drift [129] . Ac-
ording to this phenomenon, users tend to have different interests
hen they travel out of their hometown. To this aim, the authors
ropose a model to simulate the decision-making process of users’
heck-in behaviors, both in hometown and out-of-town areas, us-
ng both geographical clustering and topic modeling (i.e., Latent
irichlet Allocation – LDA [130] ) on users’ check-ins and reviews.
Furthermore, other works explore the influence of social aspects
n location-based recommendations. Assuming that friends share
any more preferences than strangers, Ye et al. [131] limit the
hoice of neighbors to the user’s friends only. They extend their
ork in [132] by selecting candidate neighbors taking into account
he user preferences extracted from check-ins histories and so-
ial connections, but also the geographical distance between users
nd candidate places. They also perform a comparative evalua-
ion among different CF and random walk-based approaches, prov-
ng that the geographical factor has more impact on the accuracy
f the recommendations than personal social relations. However,
ocial links better address the cold-start problem because social
riends may provide potentially relevant POIs to new users without
ocation histories. Therefore, the combination of different aspects,
uch as geographical influence, social relations and user prefer-
nces, ensures to obtain the highest performance in recommending
ndividual locations [133] .
As previously described, SRS can provides people/friend recom-
endations mainly exploiting the underlying social structure and
ser interaction patterns [74,77,134] . However, it has been proven
hat location has significant correlation with users’ social behaviors
n their real life, such as friendship relationships [135,136] . There-
ore, LBSNs can provide also a new way to make people/friend rec-
mmendations taking into account the users’ location histories. For
xample, Zheng et al. [137] and Ying et al. [138] focus on finding
opular users in LBSNs. They consider as “popular” the users with
ore knowledge about the locations (e.g., those who made sev-
ral check-ins) and they assume that this kind of users are able
o provide high quality location recommendations. A large body
f research (e.g., [139–143] ) leverages on users’ location histories
o improve the effectiveness of friend recommendations. Here, the
ain idea is that the history of visited places can reveal prefer-
nces, and thus people with similar location histories may have
M.G. Campana, F. Delmastro / Online Social Networks and Media 3–4 (2017) 75–97 89
s
e
u
i
r
r
t
b
a
r
t
l
t
s
t
t
s
o
a
t
i
w
[
s
t
a
t
m
o
s
p
c
t
t
i
g
v
t
[
o
t
r
a
S
t
e
T
i
i
m
5
s
v
I
t
t
t
i
f
m
(
i
t
e
o
m
a
r
b
c
t
i
e
m
m
t
I
w
t
o
t
a
l
t
s
m
s
h
d
o
n
t
f
n
m
i
m
5
c
a
D
r
h
C
t
C
c
o
5
w
S
d
b
s
a
t
t
imilar preferences and may become friends more likely. Scellato
t al. [143] propose a supervised learning framework that exploits
sers’ visited locations to predict new links among place-friends ,
.e., users who visit the same places, and friends-of-friends . Their
esults show how the inclusion of information about places, and
elated user activities, offers high link prediction probability for
he friend recommendation task. Another example is represented
y Xiao et al. [140] , which estimates the similarity between users
ccording to the semantic of their visited locations (e.g., restau-
ants or cinemas). This allows the system to recommend connec-
ions between users who have different geographic behaviors (e.g.,
iving in different cities), but share similar semantic behaviors (e.g.,
hey frequently visit the same types of locations).
Leveraging on the data shared by users in LBSNs, it is also pos-
ible to extract useful knowledge about locations and user activi-
ies. For instance, users may be interested in knowing which ac-
ivities (e.g., dining, shopping, watching movies/shows, enjoying
ports/exercises) can be practiced in a given place. Activity rec-
mmendations can be performed by exploiting both the meta-data
ssociated with the locations (e.g., tags) [144] and the users’ loca-
ion histories [145] . On the other hand, other works provide activ-
ty recommendations by exploiting information from all the users
ith collaborative-based approaches. For instance, Zheng et al.
146] and Symeonidis et al. [147] propose the use of a 3-order ten-
or to characterize each user check-in operation. In this case, the
ensor includes the current user’s location, the performed activity
nd a rating on that activity. They use factorization methods on
hese tensors in order to provide location-specific activity recom-
endations.
The last type of recommendations that can benefit from the use
f geographical information is the content recommendation, with
pecific reference to social media contents. In this case, the system
rovides users with suggestions of photos, videos, or other web
ontent they might like, which have been shared by other users
hrough a OSN. Several works exploit the histories of users loca-
ion to improve the quality of social media recommendations. For
nstance, Kawakubo and Yanai [148] proposes a picture ranking al-
orithm that exploits the locations in order to improve the rele-
ance of search results. Silva and Martins [149] presents a method
o suggest the most relevant tags for georeferenced photos, and
150] builds a location-based RS aimed at increasing the diversity
f recommendations, while mitigating popularity bias of web con-
ents. In a recent work, Pálovics et al. [151] address the problem of
ecommending highly volatile items, as locations or events char-
cterized by strict time constraints (e.g., a concert or an expo).
pecifically, they use both online machine learning and matrix fac-
orization techniques to recommend to their users the most rel-
vant hashtags for a given geo-referenced tweet at a given time.
heir solution performed much better than online matrix factor-
zation [152] and content-based [153] methods, which justifies the
mportance of temporal and geographic information in the social-
edia recommendation task.
. Recommender Systems for Mobile Social Networks
In Mobile Social Networks (MSN), typically, a mobile device
hould be able to autonomously share information with other de-
ices in proximity in order to discover useful contents for its user.
n this context, RS represent a useful tool to improve the con-
ent dissemination in MSN, by proactively suggesting to the users
he information discovered in the nearby. However, given the dis-
ributed nature of MSN, the solutions proposed for OSN (discussed
n Section 4 ) cannot be simply extended to this new scenario. In
act, RS for OSN generally rely on standard client/server centralized
odels, where the recommendation engine runs on the server side
or a cloud-based infrastructure) and processes the requests com-
ng from fixed and mobile clients. In addition, the recommendation
ask in MSN significantly differs from that defined for the online
nvironment. As discussed in Section 2 , RS for OSN typically focus
n learning a predicting model for missing values in the ratings
atrix. This matrix can be seen as a systems’ global knowledge
bout both the available items and users’ preferences. Thus, the
ecommendation engine learns a single global model which will
e used to perform the recommendations for all its users. On the
ontrary, in MSN each device may be aware of just a small part of
he global information. This local knowledge is initially related to
nformation about the local user only, then it grows up with those
xchanged with encountered users and devices, through D2D com-
unications. Therefore, in MSN each device should learn a recom-
endation model focused on the preferences of its local user.
The first approaches proposed in the literature for RS in dis-
ributed environments refer to the Peer-to-Peer (P2P) paradigm.
n a P2P architecture, RS are fully distributed over several nodes,
hich act as client and server at the same time [154] . The solu-
ions proposed in this context (e.g., [155–157] ) are typically based
n Neighborhood CF. Specifically, the user–item ratings are main-
ained in a distributed way on the P2P infrastructure and, when
peer has to make a prediction for its local user, it exploits P2P
ookup methods to find and retrieve the relevant ratings necessary
o locally calculate the rating for the target item.
By shifting the recommendation process to the client side, P2P
olutions are able to alleviate some critical points of centralized
ethods; for example, the possible bottleneck represented by the
erver, or security and privacy issues that can arise when data is
andled by a single organization [158,159] . However, by using stan-
ard P2P systems, peers should be able to communicate to each
ther by using infrastructured networks (i.e., accessing the Inter-
et), in order to have a constant access to the entire knowledge of
he network. This is not always possible in MSN and, in the last
ew years, some RS explicitly tailored for MSN and opportunistic
etworks have been presented in the literature.
We classify the proposed solutions based on the RS
ethod used: (i) CF, and (ii) tag-based, previously described
n Sections 3.1 and 4.2 , respectively. Hereafter we present the
ain solutions belonging to the two categories.
.1. Collaborative Filtering
The first set of RS proposed in the field of MSN uses CF. In this
ase, the recommendation task is performed on each local device,
nd it is based on the users’ ratings exchanged among nodes via
2D communications.
The proposed solutions differ in two main aspects: (i) the met-
ics used to measure the similarity between users, and/or (ii) the
euristics implemented to reduce the computational complexity of
F in a mobile environment. The simplest similarity measures refer
o the rating-based metrics used in standard CF (e.g., the Pearson
oefficient discussed in Eq. (9) ), while other solutions take into ac-
ount additional context information to further improve the quality
f the recommendations.
.1.1. OppCF
De Spindler et al. [160] are among the first in the literature
ho investigated the use of RS in opportunistic environments.
pecifically, they use a distributed user-based CF approach to pre-
ict ratings based on shared information among co-located mo-
ile devices. They present a preliminary idea of “social context ”
hared among mobile devices during their opportunistic contacts
nd aimed at improving the recommendation process. In this case,
he social context is identified by the co-location of users related
o a specific event (e.g., attending a conference or a concert). This
90 M.G. Campana, F. Delmastro / Online Social Networks and Media 3–4 (2017) 75–97
a
t
u
a
t
m
c
f
R
v
f
r
5
o
r
s
s
s
e
t
l
u
v
t
t
u
t
r
r
o
g
a
i
c
t
v
v
p
t
t
m
i
t
fi
p
o
p
o
i
t
f
c
e
t
S
s
i
a
p
condition is the basic assumption of the authors to identify simi-
lar users, as those who are physically co-located in a specific time
interval. The authors do not consider the possibility to exploit also
the occasional co-location of users as potential source of interest-
ing contents. This pioneering solution has never been evaluated,
thus its accuracy cannot be assessed.
5.1.2. MobHinter
MobHinter [161] presents two main contributions: (i) the defi-
nition of Affinity networks aimed at identifying similar users, and
(ii) an epidemic protocol designed to distribute and exchange
users’ ratings among mobile devices in ad-hoc networks. The au-
thors focus on the concept of affinity on ratings, by calculating the
percentage of similar ratings provided by two users on a common
set of items, and by assigning a threshold value to identify sim-
ilar users. Therefore, to apply CF method based on users’ affini-
ties, MobHinter proposes an epidemic protocol to exchange a set
of ratings information through ad hoc communications: (i) the list
of ratings of the local node, (ii) the list of ratings of its current
neighbors and (iii) the ratings of the nodes encountered in the
past. Then, each node is able to locally compute its affinity net-
work, and to apply standard CF methods to predict missing rat-
ings and provide the recommendation. The authors proposed also
optimized versions of the protocol, by exchanging only the list of
ratings of the local node or by excluding those of the nodes en-
countered in the past.
Authors evaluate their proposal using the MovieLens dataset, by
simulating a fully connected scenario (to estimate the optimal be-
havior) and an ad hoc network scenario based on a random mobil-
ity model. In the former scenario, they compared their model with
other user-based CF methods, evaluating the prediction accuracy
(i.e., MAE and RMSE). Then, they performed a simulation to evalu-
ate the number of random meetings needed on average by users to
approximate the reference prediction accuracy obtained in the fully
connected scenario. The authors showed that by exchanging all the
ratings information known by each user, the system approximates
the reference accuracy with a limited number of meetings. How-
ever, the use of a random mobility model is not realistic to simu-
late an opportunistic networking scenario as largely demonstrated
in the literature [162] .
5.1.3. DiffeRS
Del Prete and Capra [163] propose an interesting approach to
reduce the complexity of CF in MSNs. They propose diffeRS [163] ,
a decentralized RS aimed at locally classify the user as mass-like
minded or individual , based on the increasing knowledge about the
other encountered users and their profiles. In fact, also in this case,
mobile devices exchange the users profile through D2D communi-
cations (by exploiting Bluetooth technology, in this case). For each
recommendation task, diffeRS evaluates the average deviation of
the local user’s profile with respect to the preferences expressed by
the local community as a whole (i.e., the average ratings). If the lo-
cal user is identified as mass-like minded , the recommendations are
simply based on the average of the community preferences, thus
reducing at the minimum the computation on the mobile device.
On the other hand, if the local user is an individual , a user-based
CF approach is applied, considering only other individual users that
differ from the community’s preferences in the same proportion of
the local user. To this aim, the system selects only those neighbors
that share with the local user at least one rated item, and it com-
putes the difference among their co-ratings to define a measure
of their similarity with respect to their deviation from the average
of the community preferences. In this way, diffeRS further reduces
the sparsity of the rating matrix used in CF and, consequently, the
recommendation complexity for both user profiles, by exchanging
a limited amount of information during the opportunistic contacts.
The authors evaluate diffeRS both in a fully connected scenario
nd in an opportunistic one by using two datasets: MovieLens
o model users’ ratings, and MIT Reality Mining [164] to model
sers’ mobility. In the fully connected scenario, the authors evalu-
te the prediction accuracy in terms of MAE and they demonstrate
hat the proposed correlation measure, divided between mass-like
inded and individual, performs better than the standard Pearson
orrelation.
In the opportunistic scenario, they randomly map100 users
rom the MovieLens ratings dataset to the 100 users of the MIT
eality Mining dataset. Predictions are computed at regular inter-
als of time, comparing accuracy (i.e., MAE) and coverage of dif-
eRS with respect to CF based on Pearson correlation. The obtained
esults show that diffeRS always outperforms the user-based CF.
.1.4. LocPref
In a recent work, Zhao et al. [165] propose the use of CF to rec-
mmend privacy preferences depending on the user location. They
efer to a Location Sharing Service (LSS) scenario, where users can
hare their current location using the “check-in” functionality of
ocial network applications. The declaration of a user to be in some
pecific location (e.g., a clinic or religious sites) may be consid-
red as a sensitive information, and people usually may not wish
o share them with the others [166] . LSS can autonomously share
ocation information, following the general privacy settings of the
ser. LocPref is aimed at recommending the user with specific pri-
acy settings for sharing her current location through the LSS. To
his aim, it leverages on the assumption that people who visited
he same location can share similar privacy preferences. Therefore,
sers’ mobile devices can exchange location-privacy preferences
hrough opportunistic communications and locally compute their
ecommendations.
The authors also propose a simple but elegant mechanism to
educe the vulnerability of CF to shilling attack [167] : the tentative
f malicious users (i.e., attackers) to bias the recommendations by
iving false information and/or ratings [168] . The attackers can cre-
te fake profiles (also called shill profiles ) in order to inject mod-
fied data into the recommendation process. For instance, in the
ontext of LSSs, a business owner may force the customers to share
heir location when they visit her shops in order to increment their
isibility on social media. In the same way, she may want to pre-
ent the customers’ check-ins at competitors’ shops to reduce their
opularity.
In order to prevent shilling attacks, authors propose a repu-
ation scheme based on the encounter frequency of nodes. Every
ime a node receives a user’s profile from another node, it incre-
ents a counter associated with that specific profile and considers
t as the reputation level of the encountered user. Once the sys-
em needs to perform a recommendation, it uses only those pro-
les whose counter is greater than the average reputation of all the
rofiles in the local cache (i.e., the average number of encounters
f all nodes in the cache). An attacker usually creates multiple fake
rofiles, but in an opportunistic environment, it can share just one
f them during a contact. For this reason, this simple mechanism
s able to associate a lower reputation with the shill profiles than
hose calculated for real ones.
To evaluate the proposed solution, locPref implements two dif-
erent schemes for the opportunistic data exchange: (i) nodes ex-
hange only the local user profile (in terms of privacy settings) at
ach opportunistic contact ( D-Ind ); and (ii) nodes exchange both
he local profile and those previously received by the others ( D-
et ). In order to obtain a reference evaluation metric, the authors
imulate the system in a fully connected environment and then
n a mobile scenario. They exploit the st_andrews/locshare dataset
vailable in CRAWDAD [169] , which contains the location-privacy
references of 40 users collected in the city of St Andrews (UK),
M.G. Campana, F. Delmastro / Online Social Networks and Media 3–4 (2017) 75–97 91
a
r
S
s
a
c
m
o
R
5
f
r
a
o
n
j
b
t
M
5
a
i
H
p
t
n
c
M
o
t
n
o
e
t
t
t
t
t
l
t
d
t
fi
t
f
t
t
c
s
f
n
r
s
t
r
p
p
a
m
t
t
e
5
c
I
a
t
r
o
m
a
m
p
G
t
k
t
i
e
n
c
f
t
[
n
r
s
p
(
r
c
(
l
5
a
r
m
b
w
u
s
t
e
n
c
t
a
P
a
p
u
t
t
t
7 https://developer.twitter.com/en/docs/tweets/filter-realtime
nd they simulate the user mobility patterns based on St Andrews
oad map and selected POIs in the city.
Simulations results show a comparison between D-Ind and D-
et versions of locPref, highlighting the time needed to reach the
ame accuracy of the fully connected scenario (comparable with
centralized solution) and the advantages of exploiting the ex-
hange of additional profile information to compute the recom-
endations. In addition, the authors demonstrate the efficiency
f their reputation-based mechanism, which allows to prevent the
ecommender System from being abused by shilling attacks.
.2. Exploiting tags
As we described in Section 4.2 , exploiting user-defined tags (i.e.,
olksonomies) is an effective method to improve the quality of the
ecommendations. Tags can be also used in MSN to further char-
cterize contents even from a semantic point of view. However, in
rder to implement tag-based MSN solutions, we need a mecha-
ism which is able to infer the relations among different tags using
ust the user’s preferences and the limited information obtained
y other nodes in proximity. In the following, we present in detail
he two solutions proposed in the literature, which exploit tags in
SNs: ICe-Habit [170] and p-PLIERS [112] .
.2.1. Information-Centric Habit (ICe-Habit)
To the best of authors’ knowledge, ICe-Habit [170] is the first
pproach proposed in the literature that exploits folksonomies to
mprove the performance of a content dissemination protocol (i.e.,
abit [171] ) for opportunistic networks. In ICe-Habit, each user’s
rofile is characterized by a vector of tags used in the past by
he user to describe her contents. User profiles are exchanged by
odes during opportunistic contacts. In this way, each node can lo-
ally build a tag co-occurrence matrix M . Each entry of the matrix,
[ i, j ] represents the number of items characterized by the co-
ccurrence of tags ( t i , t j ). Therefore, M is used to select contents
o be forwarded to potentially interested users in opportunistic
etworks.
Whenever a user creates a new content, she associates a set
f tags T = { t 1 , . . . , t n } with it. Then, the system performs the tag-
xpansion procedure to further enrich the set of tags T . Specifically,
he system queries the matrix M to extract the k tags (i.e., T k ) with
he highest value of co-occurrence with those specified in T . Then,
he expansion set of tags T ′ = T ∪ T k is associated to the new con-
ent. Finally, the item is opportunistically forwarded to the poten-
ially interested users, i.e., each user x whose profile T x contains at
east one of the tags in T ′ . The authors evaluate ICe-Habit, via simulation, in an oppor-
unistic scenario. To this aim, they use the MIT Reality Mining
atasets, which contains the contact traces of 96 nodes. In addi-
ion, the MovieLens dataset has been used to create a user’s pro-
le for each node. Specifically, each profile contains the set of tags
he user has created to tag movies in the dataset. Furthermore,
or each < user i , tag j > pair, the MovieLens dataset contains also
he timestamp information in which the tag j has been created by
he user i . The authors use the timestamp information to map the
reation of new items on the simulated time. In order to mea-
ure the performance of the proposed solution, authors use the
ollowing approach: when a node creates a new content, it does
ot include all the tags it would normally associate with it, but it
andomly drops 50% of the item’s tags. In this way, authors mea-
ure the Tag Recovery and Destination Recovery , which are, respec-
ively, the percentage of the dropped tags and correct destination
eached by each message that could be recovered by the tag ex-
ansion method.
Although the results demonstrate that the tag expansion im-
roves the content dissemination’s performance in a MSN, they
lso prove that this approach is not able to correctly infer the se-
antic relations between different tags. In fact, considering just
he tags’ co-occurrences, nodes could receive more items than
hose really interesting for them because users may not be inter-
sted in the topics represented by the expanded tags.
.2.2. Pervasive PLIERS (p-PLIERS)
Recently, we proposed a novel framework for the content dis-
overy and evaluation in MSN called pervasive PopuLarity-based
tEm Recommender System ( p-PLIERS ) [112] . It can be used both
s standalone RS for MSN, and as a useful support tool for con-
ent dissemination or routing protocols in fully distributed envi-
onments (e.g., opportunistic networks).
Differently from ICe-Habit [170] , p-PLIERS does not leverage just
n the tags’ co-occurrences. It exploits all the folksonomy’s infor-
ation, i.e., the relations among users, shared or created items,
nd tags used to semantically describe the contents. This infor-
ation is locally maintained by each node, and modeled as a tri-
artite users–items–tags graph, which is called Local Knowledge
raph ( LKG ). The LKG of each node merges the information about
he items created or downloaded by its local user, with the local
nowledge of other encountered nodes, obtained through oppor-
unistic communications. Specifically, when a node creates a new
tem, it updates its LKG by inserting the relation between its user
ntity, the generated item and the related tags. Similarly, when a
ode encounters another node in the network, the two nodes ex-
hange their LKGs, and locally integrate them with the received in-
ormation.
Once each node has updated its local LKG with the informa-
ion coming from the other node, p-PLIERS implements PLIERS
111] (described in Section 4.2 ) to evaluate the relevance of the
ew available items with respect to the interests of its local user.
As a first set of experiments, we evaluate the accuracy of PLIERS
ecommendations in a centralized scenario with respect to other
olutions proposed for MSN: the Tag Expansion mechanism ex-
loited by ICe-Habit [170] and user-based CF used in similar works
e.g., MobHinter [161] or diffeRs [163] ). To this aim, we extract a
eal dataset from Twitter using the Twitter Streaming API 7 . Specifi-
ally, we have downloaded the tweets generated during a big event
i.e., the World Food Day at Expo 2015) in the urban area of Mi-
an, and we have built a tripartite graph composed by more than
0 0 0 users, 2946 tweets, and 3202 hashtags. This represents a re-
listic folksonomy user–item–tag graph of online tagged contents
elated to a popular event. To assess the accuracy of the recom-
endations, we perform a link prediction task on a tripartite graph
y removing one link from each user connected at least to 5 items
ith popularity greater than one (i.e., the number of connected
sers). Then, we calculate the performance of PLIERS, Tag Expan-
ion and user-based CF in terms of precision and recall, based on
he number of links that are included in the recommendations of
ach algorithm (i.e., the “recovered links”). In the centralized sce-
ario, PLIERS outperforms the two reference algorithms for both
onsidered measures. Specifically, it reaches a precision score up
o eight times higher than that obtained by Tag Expansion and
score 40% higher than CF. With regard to the Recall measure,
LIERS obtains a score around 50% higher than Tag Expansion and
round 30% higher than CF. These results indicate that PLIERS, ex-
loiting all the information contained in the folksonomy graph (i.e.,
ser–item–tag relationships), is able to obtain better results than
he other solutions proposed for MSN, which consider only part of
he user–item or item–tag relationships.
In order to validate the overall framework, we have simulated
hree realistic scenarios: a big event (Expo 2015), a conference
92 M.G. Campana, F. Delmastro / Online Social Networks and Media 3–4 (2017) 75–97
. . .
Fig. 13. Map of Expo 2015 area with the position of five of the simulated communities [112] .
a
r
c
g
r
e
c
t
w
e
a
a
t
t
i
i
o
a
m
6
p
p
p
d
r
e
b
i
l
t
t
i
a
i
i
p
O
t
m
f
t
c
venue (ACM KDD 2015), and a working day in the Helsinki city
center. In this case, we investigate the efficiency of p-PLIERS in a
fully decentralized environment, where the contents are dynami-
cally generated over time by nodes. In particular, we measure the
similarity of the recommendations generated by the framework on
the nodes’ LKGs and on the global knowledge graph (GKG), that is,
the union of all the LKGs at each time-step of the simulations. To
this aim, we use both real and synthetic mobility traces, and we
build three different datasets from Twitter for the nodes’ content
generation during the simulations. Specifically, for the Expo 2015
scenario, we simulate different sets of nodes (i.e., 250, 500, and
900 nodes) moving in the Expo area of Milan using the HCMM
[172] human mobility model. Using this model, we are able to sim-
ulate the presence of different communities, which fits well with
the reference scenario, as the Expo area was divided into several
pavilions. Fig. 13 depicts the simulated area in the Expo scenario
with the communities of nodes used in HCMM. In addition, we
use the dataset related to the World Food Day in Expo 2015 as
content (i.e., tweets) generation pattern of the nodes during the
simulations.
As a second dynamic scenario, we consider a school campus
during a conference event, where people stay most of the time
within rooms, but they regularly gather at breaks (e.g., coffee or
lunch breaks). Therefore, we use real contact traces represent-
ing the physical interactions of a group of students, professors,
and staff of an American high school during a typical school day
(i.e., the high resolution Human Contact Network – HCN [173] ). As
American high schools are not organized into classes as in Europe,
but rather around study tracks, and students are free to decide
which lectures to attend, we think that their movements can be as-
similated to those of people attending a large conference. For this
reason, we have downloaded the tweets generated during a large
computer science conference (i.e., the 21st ACM SIGKDD Confer-
ence on Knowledge Discovery and Data Mining), and we use this
data to simulate the nodes’ content generation during the simula-
tion.
Finally, in order to simulate the use of p-PLIERS on a larger
scale than the previous scenarios, we extract the contact traces of
a typical working day in Helsinki using a realistic human mobility
model highly customized on the considered area. Specifically, we
have used the ONE [174] simulator, which implements the Working
Day Mobility Model ( WDMM ) [175] and the map of the city cen-
ter of Helsinki. Then, we have downloaded the tweets generated
within the same geographic area to use them as content generated
by nodes during the simulation.
In the best case, i.e., the conference scenario, the average simi-
larity reaches 80% already in two hours simulated time, over eight
hours duration. In the worst case, i.e., Helsinki scenario, the aver-
t
ge similarity reaches 80% at the end of the simulation, due to the
educed number of contacts during the working day.
This implies that, in dense scenarios, nodes require just few
ontacts to well approximate the global knowledge with their local
raphs, demonostrating that p-PLIERS is able to provide effective
ecommendations, comparable to those achievable if global knowl-
dge is accessible. For the urban scenario of Helsinki, in which the
onsidered area is much larger than that of the other scenarios and
he density of nodes is lower, the results indicate that a larger time
indow is required for a good approximation of the global knowl-
dge about contents in the network. With a view to smart cities,
possible solution to improve the diffusion of knowledge and the
ccuracy of p-PLIERS might be based on exploiting also the public
ransportation system nodes (e.g., buses, trams, or taxis) as addi-
ional information carriers.
In Table 1 we summarize and compare the solutions presented
n the previous sections specifying, for each of them, the method
mplemented for recommendation, the context information used to
ptimize the process, and the details of their performance evalu-
tion in terms of (i) dataset, (ii) mobility traces and (iii) perfor-
ance metrics.
. Concluding remarks and open challenges
In this work, we have presented a survey of the main RS pro-
osed in the literature for Online and Mobile Social Networks, with
articular attention to the use of social context information to im-
rove the recommendation process. We describe advantages and
rawbacks of standard recommendation techniques in these envi-
onments and we highlight RS challenges in the fully distributed
nvironment of Mobile Social Networks. Several solutions have
een proposed for OSN, while the study of efficient RS for MSN
s still in its infancy and presents several open research challenges,
ike the use of additional context information, specifically related
o the mobile environment, to further optimize the recommenda-
ion process.
Due to their sensing capabilities and the constant presence
n human daily life, modern personal devices (e.g., smartphones
nd tablets) represent the bridge between the cyber and phys-
cal worlds, characterizing the situation in which the user is
nvolved during everyday life. Therefore, mining the information
rovided by these devices, and properly combining them with
SN data, can provide a more accurate model of the user’s con-
ext and preferences. For example, the social context of the user
ay be defined combining the virtual social relationships extracted
rom OSNs with the physical contacts among the devices and
he information contained in the user’s personal device (e.g., the
ontacts saved in the address book, call logs, and messages). In
his case, the user’s preferences and needs can be inferred from
M.G. Campana, F. Delmastro / Online Social Networks and Media 3–4 (2017) 75–97 93
Table 1
RS for MSN.
Algorithm RS method Context info Dataset Evaluation mobility scenario Metrics
OppCF [160] User-based CF
Location
Time
Ratings
Mobhinter [161] User-based CF Ratings MovieLens Random
RMSE
MAE
DiffeRS [163] User-based CF + heuristic Ratings MovieLens MIT Reality MAE
Mining [164] Coverage
LocPref [165] User-based CF + heuristic
Location
LocShare Map-based
Custom
Time Accuracy metric
Ratings
ICe-Habit [170] Tags correlations Tags MovieLens MIT Reality
Recall Mining [164]
p-PLIERS [112] Graph- and diffusion-based
Users
HCMM [172] Precision
Items HCN [173] Recall
Tags WDMM [175]
h
a
a
a
a
s
b
a
c
p
t
a
v
t
t
t
e
t
q
e
R
t
b
i
a
p
e
t
o
o
i
t
A
p
R
U
F
R
R
eterogeneous sources of data, e.g., the web browser history, the
ctions performed in OSNs, the generated contents both in OSN
nd MSN, the visited locations, and the set of most used mobile
pps.
The availability of this heterogeneous information represents
crucial aspect in building and evaluating new recommending
olutions. However, in order to validate and evaluate the possi-
le solutions in realistic environments, it is necessary to gener-
te new datasets, collecting all the contextual information that can
haracherizes characterise the mobile environment and can be ex-
loited in MSN. Collecting new datasets is a very time consuming
ask, requiring the implementation of a real prototype application
nd the involvement of a high number of mobile users in real en-
ironments. A possible solution is represented by the use of syn-
hetic (but realistic) context data generator (e.g., [176,177] ) but, at
he moment, none of these solutions is able to integrate realistic
emporal settings in the data generation process. Since user’s pref-
rences are highly dynamic, especially in a mobile environment,
he use of an obsolete preference model could negatively affect the
uality of the recommendations.
In addition, several frameworks have been recently proposed to
valuate RS performance in OSN, e.g., LibRec [178] , LensKit [179] ,
ankSys [180] , and CARSKit [181] . They provide an implementa-
ion of some RS methods (e.g., CF, MF) and performance metrics,
ut they are developed exclusively for centralized RS, not consider-
ng the unique characteristics of MSNs (e.g., the limited knowledge
vailable in each device). Therefore, in order to perform a fair com-
arison among different RS for MSN, we should define a common
valuation framework with the following three main characteris-
ics:
• Mobility : RS should be evaluated in realistic mobile scenarios.
Therefore, the evaluation framework should include datasets of
real mobility traces and human mobility models.
• Data : as we previously pointed out, RS for MSN should char-
acterize the user’s context with several heterogeneous data. To
this aim, the framework should also include proper datasets or
context data generators. In addition, given the extremely dy-
namic nature of the MSN environment, during the simulation
nodes should be able to generate data over time.
• Evaluation : the framework should include the main evaluation
metrics proposed in the literature for the two main recommen-
dation tasks (i.e., prediction and ranking). Furthermore, due to
the typically limited duration of the opportunistic contacts, it is
also needed to consider the time complexity of the RS in MSN.
In fact, if the RS is not able to provide recommendations in a
limited time window, a node may miss the opportunity to ex-
change useful data with another one in proximity.
All these aspects represent a wide research area, including not
nly the definition of new RS algorithms, but also the creation
f experimental testbeds including new opportunistic network-
ng protocols, opportunistic sensing features and context reasoning
ools.
cknowledgments
This work was carried out in the framework of the INTESA
roject (CUP CIPE D78I160 0 0 010 0 08), co-funded by the Tuscany
egion (Italy) under the Regional Implementation Programme for
nderutilized Areas Fund (PAR FAS 2007–2013) and the Research
acilitation Fund (FAR) of the Ministry of Education, University and
esearch (MIUR).
eferences
[1] F. Ricci , Mobile recommender systems, Inf. Technol. Tour. 12 (3) (2010)
205–231 . [2] I. Guy, D. Carmel, Social recommender systems, in: Proceedings of the Twen-
tieth International Conference Companion on World Wide Web, WWW ’11,
ACM, New York, NY, USA, 2011, pp. 283–284, doi: 10.1145/1963192.1963312 . [3] M. Conti, S. Giordano, M. May, A. Passarella, From opportunistic networks to
opportunistic computing, IEEE Commun. Mag. 48 (9) (2010) 126–139, doi: 10.1109/MCOM.2010.5560597 .
[4] M. Conti , A. Passarella , S.K. Das , The Internet of People (IoP): A new wave inpervasive mobile computing, Pervasive Mob. Comput. 41 (2017) 1–27 .
[5] Y. Shi , M. Larson , A. Hanjalic , Collaborative filtering beyond the user-item ma-
trix: a survey of the state of the art and future challenges, ACM Comput. Surv.(CSUR) 47 (1) (2014) 3 .
[6] Y.-J. Park , A. Tuzhilin , The long tail of recommender systems and how toleverage it, in: Proceedings of the 2008 ACM Conference on Recommender
Systems, ACM, 2008, pp. 11–18 . [7] T. Mahmood , F. Ricci , Learning and adaptivity in interactive recommender
systems, in: Proceedings of the Ninth International Conference on Electronic
Commerce, ACM, 2007, pp. 75–84 . [8] T. Mahmood , F. Ricci , Improving recommender systems with adaptive con-
versational strategies, in: Proceedings of the Twentieth ACM Conference onHypertext and Hypermedia, ACM, 2009, pp. 73–82 .
[9] N. Taghipour , A. Kardan , S.S. Ghidary , Usage-based web recommendations: areinforcement learning approach, in: Proceedings of the 2007 ACM Confer-
ence on Recommender Systems, ACM, 2007, pp. 113–120 . [10] C.J. Willmott , K. Matsuura , Advantages of the mean absolute error (MAE) over
the root mean square error (RMSE) in assessing average model performance,
Clim. Res. 30 (1) (2005) 79–82 . [11] L. Baltrunas, T. Makcinskas, F. Ricci, Group recommendations with rank aggre-
gation and collaborative filtering, in: Proceedings of the Fourth ACM Confer-ence on Recommender Systems, RecSys ’10, ACM, New York, NY, USA, 2010,
pp. 119–126, doi: 10.1145/1864708.1864733 .
94 M.G. Campana, F. Delmastro / Online Social Networks and Media 3–4 (2017) 75–97
[12] S. Balakrishnan , S. Chopra , Collaborative ranking, in: Proceedings of the FifthACM International Conference on Web Search and Data Mining, ACM, 2012,
pp. 143–152 . [13] J.L. Herlocker , J.A. Konstan , L.G. Terveen , J.T. Riedl , Evaluating collaborative fil-
tering recommender systems, ACM Trans. Inf. Syst. (TOIS) 22 (1) (2004) 5–53 .[14] A. Gunawardana , G. Shani , Evaluating recommender systems, in: Recom-
mender Systems Handbook, Springer, 2015, pp. 265–308 . [15] F. Ricci , L. Rokach , B. Shapira , Introduction to recommender systems hand-
book, in: Recommender Systems Handbook, Springer, 2011, pp. 1–35 .
[16] J. Tang , X. Hu , H. Liu , Social recommendation: a review, Soc. Netw. Anal. Min.3 (4) (2013) 1113–1133 .
[17] X. Ning , C. Desrosiers , G. Karypis , A comprehensive survey of neighbor-hood-based recommendation methods, in: Recommender Systems Handbook,
Springer, 2015, pp. 37–76 . [18] J. Herlocker , J.A. Konstan , J. Riedl , An empirical analysis of design choices in
neighborhood-based collaborative filtering algorithms, Inf. Retr. 5 (4) (2002)
287–310 . [19] C.C. Aggarwal , Recommender Systems, Springer, 2016 .
[20] B. Mobasher , H. Dai , T. Luo , M. Nakagawa , Effective personalization based onassociation rule discovery from web usage data, in: Proceedings of the Third
International Workshop on Web Information and Data Management, ACM,2001, pp. 9–15 .
[21] M.-L. Shyu , C. Haruechaiyasak , S.-C. Chen , N. Zhao , Collaborative filtering by
mining association rules from user access sequences, in: Proceedings of the2005 International Workshop on Challenges in Web Information Retrieval and
Integration, WIRI’05, IEEE, 2005, pp. 128–135 . [22] K. Miyahara , M. Pazzani , Collaborative filtering with the simple Bayesian clas-
sifier, Proceedings of the Eighteenth Pacific Rim International Conference onPRICAI 20 0 0 Topics in Artificial Intelligence, 20 0 0, pp. 679–689 .
[23] X. Su , T.M. Khoshgoftaar , Collaborative filtering for multi-class data using be-
lief nets algorithms, in: Proceedings of the Eighteenth IEEE International Con-ference on International Conference on Tools with Artificial Intelligence, IC-
TAI’06, IEEE, 2006, pp. 497–504 . [24] Z. Xia , Y. Dong , G. Xing , Support vector machines for collaborative filtering, in:
Proceedings of the Forty-fourth Annual Southeast Regional Conference, ACM,2006, pp. 169–174 .
[25] R. Salakhutdinov , A. Mnih , G. Hinton , Restricted Boltzmann machines for col-
laborative filtering, in: Proceedings of the Twenty-fourth International Con-ference on Machine Learning, ACM, 2007, pp. 791–798 .
[26] X. He , L. Liao , H. Zhang , L. Nie , X. Hu , T.-S. Chua , Neural collaborative filtering,in: Proceedings of the Twenty-sixth International Conference on World Wide
Web, International World Wide Web Conferences Steering Committee, 2017,pp. 173–182 .
[27] Y. Koren , R. Bell , C. Volinsky , Matrix factorization techniques forrecommender
systems, Computer 42 (8) (2009) 30–37 . [28] M. Pazzani , D. Billsus , Content-Based Recommendation Systems, in:
P. Brusilovsky, A. Kobsa, W. Nejdl (Eds.), The Adaptive Web: Methodsand Strategies of Web Personalization, Springer Berlin Heidelberg, Berlin,
Heidelberg, 2007, pp. 325–341 . [29] P. Lops , M. De Gemmis , G. Semeraro , Content-based recommender systems:
state of the art and trends, in: Recommender Systems Handbook, Springer,2011, pp. 73–105 .
[30] T. Miranda , M. Claypool , A. Gokhale , T. Mir , P. Murnikov , D. Netes , M. Sartin ,
Combining content-based and collaborative filters in an online newspaper, in:Proceedings of 1999 ACM SIGIR Workshop on Recommender Systems, Cite-
seer, 1999 . [31] G. Adomavicius , A. Tuzhilin , Toward the next generation of recommender sys-
tems: a survey of the state-of-the-art and possible extensions, IEEE Trans.Knowl. Data Eng. 17 (6) (2005) 734–749 .
[32] F. Yu , A. Zeng , S. Gillard , M. Medo , Network-based recommendation algo-
rithms: a review, Phys. A Stat. Mech. Appl. 452 (2016) 192–208 . [33] L. Page , S. Brin , R. Motwani , T. Winograd , The PageRank Citation Ranking:
Bringing Order to the Web., Technical report, Stanford InfoLab, 1999 . [34] T.H. Haveliwala , Topic-sensitive PageRank, in: Proceedings of the Eleventh In-
ternational Conference on World Wide Web, ACM, 2002, pp. 517–526 . [35] A. Hotho , R. Jäschke , C. Schmitz , G. Stumme , FolkRank: a ranking algorithm
for folksonomies, in: Proceedings of the 2006 Conference on Live Working
Association (LWA), vol. 1, 2006, pp. 111–114 . [36] M.-H. Tsai , C. Aggarwal , T. Huang , Ranking in heterogeneous social media, in:
Proceedings of the Seventh ACM International Conference on Web Search andData Mining, ACM, 2014, pp. 613–622 .
[37] M. Gori , A. Pucci , V. Roma , I. Siena , Itemrank: a random-walk based scor-ing algorithm for recommender engines., in: Proceedings of the 2007 In-
ternational Joint Conference on Artificial Intelligence, IJCAI, vol. 7, 2007,
pp. 2766–2771 . [38] K. Lakiotaki , N.F. Matsatsinis , A. Tsoukias , Multicriteria user modeling in rec-
ommender systems, IEEE Intell. Syst. 26 (2) (2011) 64–76 . [39] F. Fouss , A. Pirotte , J.-M. Renders , M. Saerens , Random-walk computation of
similarities between nodes of a graph with application to collaborative rec-ommendation, IEEE Trans. Knowl. Data Eng. 19 (3) (2007) 355–369 .
[40] H. Yildirim , M.S. Krishnamoorthy , A random walk method for alleviating the
sparsity problem in collaborative filtering, in: Proceedings of the 2008 ACMConference on Recommender Systems, ACM, 2008, pp. 131–138 .
[41] T. Zhou , Z. Kuscsik , J.-G. Liu , M. Medo , J.R. Wakeling , Y.-C. Zhang , Solvingthe apparent diversity-accuracy dilemma of recommender systems, Proc. Natl.
Acad. Sci. 107 (10) (2010) 4511–4515 .
[42] T. Zhou , J. Ren , M. Medo , Y.-C. Zhang , Bipartite network projection and per-sonal recommendation, Phys. Rev. E 76 (4) (2007) 046115 .
[43] J.-G. Liu , T. Zhou , Q. Guo , Information filtering via biased heat conduction,Phys. Rev. E 84 (3) (2011a) 037101 .
[44] J.-G. Liu , Q. Guo , Y.-C. Zhang , Information filtering via weighted heat conduc-tion algorithm, Phys. A Stat. Mech. Appl. 390 (12) (2011b) 2414–2420 .
[45] L. Lü, W. Liu , Information filtering via preferential diffusion, Phys. Rev. E 83(6) (2011) 066119 .
[46] T. Zhou , L.-L. Jiang , R.-Q. Su , Y.-C. Zhang , Effect of initial configuration
on network-based recommendation, Europhys. Lett. (EPL) 81 (5) (2008) 58004 .
[47] C. Liu , W.-X. Zhou , Heterogeneity in initial resource configurations improvesa network-based hybrid recommendation algorithm, Phys. A Stat. Mech. Appl.
391 (22) (2012) 5704–5711 . [48] G. Adomavicius , A. Tuzhilin , Context-aware recommender systems, in: Rec-
ommender Systems Handbook, Springer, 2011, pp. 217–253 .
[49] G. Adomavicius , R. Sankaranarayanan , S. Sen , A. Tuzhilin , Incorporating con-textual information in recommender systems using a multidimensional ap-
proach, ACM Trans. Inf. Syst. (TOIS) 23 (1) (2005) 103–145 . [50] G. Adomavicius , A. Tuzhilin , Context-aware recommender systems, in: Rec-
ommender Systems Handbook, Springer, 2015, pp. 191–226 . [51] R. Burke , Hybrid recommender systems: survey and experiments, User Model.
User-Adapt. Interact. 12 (4) (2002) 331–370 .
[52] V. Codina , F. Ricci , L. Ceccaroni , Exploiting the semantic similarity of contex-tual situations for pre-filtering recommendation, in: Proceedings of the 2013
International Conference on User Modeling, Adaptation, and Personalization,Springer, 2013, pp. 165–177 .
[53] U. Panniello , A. Tuzhilin , M. Gorgoglione , C. Palmisano , A. Pedone , Experimen-tal comparison of pre-vs. post-filtering approaches in context-aware recom-
mender systems, in: Proceedings of the Third ACM Conference on Recom-
mender Systems, ACM, 2009, pp. 265–268 . [54] A. Karatzoglou , X. Amatriain , L. Baltrunas , N. Oliver , Multiverse recommenda-
tion: n-dimensional tensor factorization for context-aware collaborative filter-ing, in: Proceedings of the Fourth ACM Conference on Recommender Systems,
ACM, 2010, pp. 79–86 . [55] S. Rendle , Factorization machines, in: Proceedings of the Tenth IEEE Interna-
tional Conference on Data Mining (ICDM), IEEE, 2010, pp. 995–10 0 0 .
[56] S. Rendle , Z. Gantner , C. Freudenthaler , L. Schmidt-Thieme , Fast context-awarerecommendations with factorization machines, in: Proceedings of the Thirty–
fourth International ACM SIGIR Conference on Research and Development inInformation Retrieval, ACM, 2011, pp. 635–644 .
[57] S. Rendle , Factorization machines with libFM, ACM Trans. Intell. Syst. Technol.(TIST) 3 (3) (2012) 57 .
[58] I. Guy , Social recommender systems, in: Recommender Systems Handbook,
Springer, 2015, pp. 511–543 . [59] R. Dunbar, V. Arnaboldi, M. Conti, A. Passarella, The structure of online so-
cial networks mirrors those in the offline world, Soc. Netw. 43 (Supple-ment C) (2015) 39–47, doi: 10.1016/j.socnet.2015.04.005 . URL: http://www.
sciencedirect.com/science/article/pii/S03788733150 0 0313 [60] V. Arnaboldi , A. Passarella , M. Conti , R.I. Dunbar , Online Social Networks –
Human Cognitive Constraints in Facebook and Twitter Personal Graphs, Com-puter Science Reviews and Trends, Elsevier, Boston, 2015 .
[61] M. McPherson , L. Smith-Lovin , J.M. Cook , Birds of a feather: homophily in
social networks, Ann. Rev. Sociol. 27 (1) (2001) 415–4 4 4 . [62] J. O’Donovan , B. Smyth , Trust in recommender systems, in: Proceedings of
the Tenth International Conference on Intelligent User Interfaces, ACM, 2005,pp. 167–174 .
[63] L. Terveen , D.W. McDonald , Social matching: a framework and researchagenda, ACM Trans. Comput.-Hum. Interact. (TOCHI) 12 (3) (2005) 401–
434 .
[64] D. Liben-Nowell , J. Kleinberg , The link-prediction problem for social networks,J. Assoc. Inf. Sci. Technol. 58 (7) (2007) 1019–1031 .
[65] V. Martínez , F. Berzal , J.-C. Cubero , A survey of link prediction in complexnetworks, ACM Comput. Surv. (CSUR) 49 (4) (2016) 69 .
[66] G. Jeh , J. Widom , SimRank: a measure of structural-context similarity, in: Pro-ceedings of the Eighth ACM SIGKDD International Conference on Knowledge
Discovery and Data Mining, ACM, 2002, pp. 538–543 .
[67] I. Guy , M. Jacovi , E. Shahar , N. Meshulam , V. Soroka , S. Farrell , Harvestingwith sonar: the value of aggregating social network information, in: Proceed-
ings of the SIGCHI Conference on Human Factors in Computing Systems, ACM,2008, pp. 1017–1026 .
[68] I. Guy , M. Jacovi , N. Meshulam , I. Ronen , E. Shahar , Public vs. private: com-paring public social network information with email, in: Proceedings of the
2008 ACM Conference on Computer Supported Cooperative Work, ACM, 2008,
pp. 393–402 . [69] I. Guy , I. Ronen , E. Wilcox , Do you know?: recommending people to invite
into your social network, in: Proceedings of the Fourteenth International Con-ference on Intelligent User Interfaces, ACM, 2009, pp. 77–86 .
[70] R. Dunbar , The social brain hypothesis, Brain 9 (10) (1998) 178–190 . [71] R.A. Hill , R.I. Dunbar , Social network size in humans, Hum. Nat. 14 (1) (2003)
53–72 .
[72] V. Arnaboldi , M. Conti , A. Passarella , F. Pezzoni , Analysis of ego network struc-ture in online social networks, in: Proceedings of the 2012 International Con-
ference on Privacy, Security, Risk and Trust (PASSAT), and 2012 InternationalConference on Social Computing (SocialCom), IEEE, 2012, pp. 31–40 .
M.G. Campana, F. Delmastro / Online Social Networks and Media 3–4 (2017) 75–97 95
[
[
[
[
[
[
[
[73] A. Passarella, R.I. Dunbar, M. Conti, F. Pezzoni, Ego network models forfuture internet social networking environments, Comput. Commun. 35
(18) (2012) 2201–2217, doi: 10.1016/j.comcom.2012.08.003 . URL: http://www.sciencedirect.com/science/article/pii/S0140366412002642
[74] A. Epasto , S. Lattanzi , V. Mirrokni , I.O. Sebe , A. Taei , S. Verma , Ego-net com-munity mining applied to friend suggestion, Proc. VLDB Endow. 9 (4) (2015)
324–335 . [75] R.N. Lichtenwalter , J.T. Lussier , N.V. Chawla , New perspectives and meth-
ods in link prediction, in: Proceedings of the Sixteenth ACM SIGKDD Inter-
national Conference on Knowledge Discovery and Data Mining, ACM, 2010,pp. 243–252 .
[76] M. Fire , L. Tenenboim , O. Lesser , R. Puzis , L. Rokach , Y. Elovici , Link predic-tion in social networks using computationally efficient topological features,
in: Proceedings of the 2011 IEEE Third International Conference on Privacy,Security, Risk and Trust (PASSAT) and 2011 IEEE Third International Confer-
ence on Social Computing (SocialCom), IEEE, 2011, pp. 73–80 .
[77] L. Backstrom , J. Leskovec , Supervised random walks: predicting and recom-mending links in social networks, in: Proceedings of the Fourth ACM Interna-
tional Conference on Web Search and Data Mining, ACM, 2011, pp. 635–644 . [78] P. Victor , C. Cornelis , M. De Cock , A. Teredesai , A comparative analysis of
trust-enhanced recommenders for controversial items., in: Proceedings ofthe 2009 International AAAI Conference on Web and Social Media (ICWSM),
2009 .
[79] J. He , W.W. Chu , A social network-based recommender system (SNRS), in:Data Mining for Social Network Data, Springer, 2010, pp. 47–74 .
[80] J. Golbeck , Generating Predictive Movie Recommendations from Trust in So-cial Networks, in: K. Stølen, W.H. Winsborough, F. Martinelli, F. Massacci
(Eds.), Trust Management: 4th International Conference, iTrust 2006, Pisa,Italy, May 16-19, 2006. Proceedings, Springer Berlin Heidelberg, Berlin, Hei-
delberg, 2006, pp. 93–104 .
[81] P. Massa , P. Avesani , Trust metrics on controversial users: balancing betweentyranny of the majority, Int. J. Semant. Web Inf. Syst. (IJSWIS) 3 (1) (2007)
39–64 . [82] M. Jamali , M. Ester , TrustWalker: a random walk model for combining
trust-based and item-based recommendation, in: Proceedings of the FifteenthACM SIGKDD International Conference on Knowledge Discovery and Data
Mining, ACM, 2009, pp. 397–406 .
[83] H. Ma , H. Yang , M.R. Lyu , I. King , SoRec: social recommendation using prob-abilistic matrix factorization, in: Proceedings of the Seventeenth ACM Con-
ference on Information and Knowledge Management, ACM, 2008, pp. 931–940 .
[84] M. Jamali , M. Ester , A matrix factorization technique with trust propagationfor recommendation in social networks, in: Proceedings of the Fourth ACM
Conference on Recommender Systems, ACM, 2010, pp. 135–142 .
[85] H. Ma , D. Zhou , C. Liu , M.R. Lyu , I. King , Recommender systems with socialregularization, in: Proceedings of the Fourth ACM International Conference
on Web Search and Data Mining, ACM, 2011, pp. 287–296 . [86] D.D. Lee , H.S. Seung , Learning the parts of objects by non-negative matrix
factorization, Nature 401 (6755) (1999) 788 . [87] R. Salakhutdinov , A. Mnih , Bayesian probabilistic matrix factorization using
Markov chain monte carlo, in: Proceedings of the Twenty-fifth InternationalConference on Machine Learning, ACM, 2008, pp. 880–887 .
[88] H. Ma , I. King , M.R. Lyu , Learning to recommend with social trust ensemble,
in: Proceedings of the Thirty-second International ACM SIGIR Conference onResearch and Development in Information Retrieval, ACM, 2009, pp. 203–210 .
[89] J.S. Breese , D. Heckerman , C. Kadie , Empirical analysis of predictive algorithmsfor collaborative filtering, in: Proceedings of the Fourteenth Conference on
Uncertainty in Artificial Intelligence, Morgan Kaufmann Publishers Inc., 1998,pp. 43–52 .
[90] Z. Sun , L. Han , W. Huang , X. Wang , X. Zeng , M. Wang , H. Yan , Recommender
systems based on social networks, J. Syst. Softw. 99 (2015) 109–119 . [91] W. Reafee , N. Salim , A. Khan , The power of implicit social relation in rating
prediction of social recommender systems, PloS One 11 (5) (2016) e0154 84 8 . [92] S. James , C.-H. Michael , The folksonomy tag cloud: when is it useful? J. Inf.
Sci. 34 (1) (2008) . [93] A. Hotho , R. Jäschke , C. Schmitz , G. Stumme , Information retrieval in folk-
sonomies: Search and ranking, in: Proceedings of the 2006 European Seman-
tic Web Conference (ESWC), vol. 4011, Springer, 2006, pp. 411–426 . [94] Z.-K. Zhang , T. Zhou , Y.-C. Zhang , Tag-aware recommender systems: a
state-of-the-art survey, J. Comput. Sci. Technol. 26 (5) (2011) 767–777 . [95] A .K. Milicevic , A . Nanopoulos , M. Ivanovic , Social tagging in recommender
systems: a survey of the state-of-the-art and possible extensions, Artif. Intell.Rev. 33 (3) (2010) 187–209 .
[96] H. Kim , H.-J. Kim , A framework for tag-aware recommender systems, Expert
Syst. Appl. 41 (8) (2014) 40 0 0–40 09 . [97] Z. Xu , Y. Fu , J. Mao , D. Su , Towards the semantic web: collaborative tag sug-
gestions, in: Proceedings of the 2006 Collaborative Web Tagging Workshop atWWW2006, Edinburgh, Scotland, 2006 .
[98] B. Sigurbjörnsson , R. van Zwol , Flickr tag recommendation based on collectiveknowledge, in: Proceedings of the Seventeenth International Conference on
World Wide Web, ACM, 2008, pp. 327–336 .
[99] C.S. Firan , W. Nejdl , R. Paiu , The benefit of using tag-based profiles, in: Pro-ceedings of the 2007 Web Conference on Latin American (LA-WEB), IEEE,
2007, pp. 32–41 .
100] I. Guy , N. Zwerdling , I. Ronen , D. Carmel , E. Uziel , Social media recommen-dation based on people and tags, in: Proceedings of the Thirty-third Interna-
tional ACM SIGIR Conference on Research and Development in InformationRetrieval, ACM, 2010, pp. 194–201 .
[101] J. Peng , D.D. Zeng , H. Zhao , F.-y. Wang , Collaborative filtering in social tag-ging systems based on joint item-tag recommendations, in: Proceedings of
the Nineteenth ACM International Conference on Information and KnowledgeManagement, ACM, 2010, pp. 809–818 .
[102] A. Shepitsen , J. Gemmell , B. Mobasher , R. Burke , Personalized recommen-
dation in social tagging systems using hierarchical clustering, in: Proceed-ings of the 2008 ACM Conference on Recommender Systems, ACM, 2008,
pp. 259–266 . [103] J. Gemmell , A. Shepitsen , B. Mobasher , R. Burke , Personalizing Navigation
in Folksonomies Using Hierarchical Tag Clustering, in: I.Y. Song, J. Eder,T.M. Nguyen (Eds.), Data Warehousing and Knowledge Discovery: 10th Inter-
national Conference, DaWaK 2008 Turin, Italy, September 2-5, 2008 Proceed-
ings, Springer Berlin Heidelberg, Berlin, Heidelberg, 2008, pp. 196–205 . 104] H. Liang , Y. Xu , Y. Li , R. Nayak , X. Tao , Connecting Users and Items with
Weighted Tags for Personalized Item Recommendations, in: Proceedings ofthe 21st ACM Conference on Hypertext and Hypermedia, ACM, {New York,
NY, USA, 2010, pp. 51–60 . [105] P. Symeonidis , A. Nanopoulos , Y. Manolopoulos , Tag recommendations based
on tensor dimensionality reduction, in: Proceedings of the 2008 ACM Confer-
ence on Recommender Systems, ACM, 2008, pp. 43–50 . 106] L. De Lathauwer , B. De Moor , J. Vandewalle , A multilinear singular value de-
composition, SIAM J. Matrix Anal. Appl. 21 (4) (20 0 0) 1253–1278 . [107] S. Rendle , L. Balby Marinho , A. Nanopoulos , L. Schmidt-Thieme , Learning op-
timal ranking with tensor factorization for tag recommendation, in: Proceed-ings of the Fifteenth ACM SIGKDD International Conference on Knowledge
Discovery and Data Mining, ACM, 2009, pp. 727–736 .
[108] D.J. Hand , R.J. Till , A simple generalisation of the area under the ROC curvefor multiple class classification problems, Mach. Learn. 45 (2) (2001) 171–186 .
[109] S. Brin , L. Page , Reprint of: the anatomy of a large-scale hypertextual websearch engine, Comput. Netw. 56 (18) (2012) 3825–3833 .
[110] Z.-K. Zhang , T. Zhou , Y.-C. Zhang , Personalized recommendation via integrateddiffusion on user–item–tag tripartite graphs, Phys. A Stat. Mech. Appl. 389 (1)
(2010) 179–186 .
[111] V. Arnaboldi , M.G. Campana , F. Delmastro , E. Pagani , Pliers: a populari-ty-based recommender system for content dissemination in online social net-
works, in: Proceedings of the Thirty-first Annual ACM Symposium on AppliedComputing, ACM, 2016, pp. 671–673 .
[112] V. Arnaboldi , M.G. Campana , F. Delmastro , E. Pagani , A personalized recom-mender system for pervasive social networks, Pervasive Mob. Comput. 36
(2017) 3–24 .
[113] Y. Zuo , J. Zeng , M. Gong , L. Jiao , Tag-aware recommender systems based ondeep neural networks, Neurocomputing 204 (2016) 51–60 .
[114] G.E. Hinton , R.R. Salakhutdinov , Reducing the dimensionality of data withneural networks, Science 313 (5786) (2006) 504–507 .
[115] Z. Xu, T. Lukasiewicz, C. Chen, Y. Miao, X. Meng, Tag-aware personalized rec-ommendation using a hybrid deep model, in: Proceedings of the Twenty-
Sixth International Joint Conference on Artificial Intelligence, IJCAI-17, 2017,pp. 3196–3202, doi: 10.24963/ijcai.2017/446 .
[116] D. Quercia , N. Lathia , F. Calabrese , G. Di Lorenzo , J. Crowcroft , Recommend-
ing social events from mobile phone location data, in: Proceedings of theTenth IEEE International Conference on Data Mining (ICDM), IEEE, 2010,
pp. 971–976 . [117] J. Bao , Y. Zheng , D. Wilkie , M. Mokbel , Recommendations in location-based
social networks: a survey, GeoInformatica 19 (3) (2015) 525–565 . [118] K. Kodama , Y. Iijima , X. Guo , Y. Ishikawa , Skyline queries based on user loca-
tions and preferences for making location-based recommendations, in: Pro-
ceedings of the 2009 International Workshop on Location Based Social Net-works, ACM, 2009, pp. 9–16 .
[119] M.-H. Park , J.-H. Hong , S.-B. Cho , Location-based recommendation system us-ing Bayesian user’s preference model in mobile devices, in: Proceedings of
the 2007 International Conference on Ubiquitous Intelligence and Computing,Springer, 2007, pp. 1130–1139 .
120] L. Ramaswamy , P. Deepak , R. Polavarapu , K. Gunasekera , D. Garg ,
K. Visweswariah , S. Kalyanaraman , Caesar: a context-aware, social recom-mender system for low-end mobile devices, in: Proceedings of the Tenth In-
ternational Conference on Mobile Data Management: Systems, Services andMiddleware, MDM’09, IEEE, 2009, pp. 338–347 .
[121] W.R. Tobler , A computer movie simulating urban growth in the detroit region,Econ. Geogr. 46 (sup1) (1970) 234–240 .
122] T. Horozov , N. Narasimhan , V. Vasudevan , Using location for personalized poi
recommendations in mobile environments, in: Proceedings of the 2006 In-ternational Symposium on Applications and the Internet, SAINT 2006, IEEE,
2006, pp. 6–pp . 123] C.-Y. Chow , J. Bao , M.F. Mokbel , Towards location-based social networking ser-
vices, in: Proceedings of the Second ACM SIGSPATIAL International Workshopon Location Based Social Networks, ACM, 2010, pp. 31–38 .
124] Y. Shi , P. Serdyukov , A. Hanjalic , M. Larson , Personalized landmark recommen-
dation based on geotags from photo sharing sites., Proceedings of the 2011Eleventh International Conference on Web and Social Media (ICWSM), vol. 11,
2011, pp. 622–625 .
96 M.G. Campana, F. Delmastro / Online Social Networks and Media 3–4 (2017) 75–97
[125] Y. Moshfeghi , B. Piwowarski , J.M. Jose , Handling data sparsity in collaborativefiltering using emotion and semantic based features, in: Proceedings of the
Thirty-fourth International ACM SIGIR Conference on Research and Develop-ment in Information Retrieval, ACM, 2011, pp. 625–634 .
[126] A. Levi , O. Mokryn , C. Diot , N. Taft , Finding a needle in a haystack of re-views: cold start context-based hotel recommender system, in: Proceedings
of the Sixth ACM Conference on Recommender Systems, ACM, 2012, pp. 115–122 .
[127] D. Yang , D. Zhang , Z. Yu , Z. Yu , Fine-grained preference-aware location search
leveraging crowdsourced digital footprints from LBSNs, in: Proceedings of the2013 ACM International Joint Conference on Pervasive and Ubiquitous Com-
puting, ACM, 2013, pp. 479–488 . [128] H. Wang , Y. Fu , Q. Wang , H. Yin , C. Du , H. Xiong , A location-sentimen-
t-aware recommender system for both home-town and out-of-town users,in: Proceedings of the Twenty-third ACM SIGKDD International Conference
on Knowledge Discovery and Data Mining, ACM, 2017, pp. 1135–1143 .
[129] H. Yin , X. Zhou , B. Cui , H. Wang , K. Zheng , Q.V.H. Nguyen , Adapting to userinterest drift for poi recommendation, IEEE Trans. Knowl. Data Eng. 28 (10)
(2016) 2566–2581 . [130] D.M. Blei , A.Y. Ng , M.I. Jordan , Latent Dirichlet allocation, J. Mach. Learn. Res.
3 (Jan) (2003) 993–1022 . [131] M. Ye , P. Yin , W.-C. Lee , Location recommendation for location-based so-
cial networks, in: Proceedings of the Eighteenth SIGSPATIAL International
Conference on Advances in Geographic Information Systems, ACM, 2010,pp. 458–461 .
[132] M. Ye , P. Yin , W.-C. Lee , D.-L. Lee , Exploiting geographical influence for col-laborative point-of-interest recommendation, in: Proceedings of the Thirty–
fourth International ACM SIGIR Conference on Research and Development inInformation Retrieval, ACM, 2011, pp. 325–334 .
[133] F. Gasparetti , Personalization and context-awareness in social local search:
state-of-the-art and future research challenges, Pervasive Mob. Comput. 38(2017) 446–473 .
[134] J. Wiese , P.G. Kelley , L.F. Cranor , L. Dabbish , J.I. Hong , J. Zimmerman , Are youclose with me? are you nearby?: investigating social groups, closeness, and
willingness to share, in: Proceedings of the Thirteenth International Confer-ence on Ubiquitous Computing, ACM, 2011, pp. 197–206 .
[135] D. Liben-Nowell , J. Novak , R. Kumar , P. Raghavan , A. Tomkins , Geographic
routing in social networks, Proc. Natl. Acad. Sci. U.S.A. 102 (33) (2005)11623–11628 .
[136] J. Cranshaw , E. Toch , J. Hong , A. Kittur , N. Sadeh , Bridging the gap be-tween physical location and online social networks, in: Proceedings of the
Twelfth ACM International Conference on Ubiquitous Computing, ACM, 2010,pp. 119–128 .
[137] Y. Zheng , L. Zhang , X. Xie , W.-Y. Ma , Mining interesting locations and travel
sequences from GPS trajectories, in: Proceedings of the Eighteenth Interna-tional Conference on World Wide Web, ACM, 2009, pp. 791–800 .
[138] J.J.-C. Ying , W.-C. Lee , M. Ye , C.-Y. Chen , V.S. Tseng , User association analysis oflocales on location based social networks, in: Proceedings of the Third ACM
SIGSPATIAL International Workshop on Location-Based Social Networks, ACM,2011, pp. 69–76 .
[139] E. Cho , S.A. Myers , J. Leskovec , Friendship and mobility: user movementin location-based social networks, in: Proceedings of the Seventeenth ACM
SIGKDD International Conference on Knowledge Discovery and Data Mining,
ACM, 2011, pp. 1082–1090 . [140] X. Xiao , Y. Zheng , Q. Luo , X. Xie , Finding similar users using category-based
location history, in: Proceedings of the Eighteenth SIGSPATIAL InternationalConference on Advances in Geographic Information Systems, ACM, 2010,
pp. 4 42–4 45 . [141] Y. Zheng , L. Zhang , Z. Ma , X. Xie , W.-Y. Ma , Recommending friends and lo-
cations based on individual location history, ACM Trans. Web (TWEB) 5 (1)
(2011) 5 . [142] X. Yu , A. Pan , L.-A. Tang , Z. Li , J. Han , Geo-friends recommendation in GP-
S-based cyber-physical social network, in: Proceedings of the 2011 Inter-national Conference on Advances in Social Networks Analysis and Mining
(ASONAM), IEEE, 2011, pp. 361–368 . [143] S. Scellato , A. Noulas , C. Mascolo , Exploiting place features in link prediction
on location-based social networks, in: Proceedings of the Seventeenth ACM
SIGKDD International Conference on Knowledge Discovery and Data Mining,ACM, 2011, pp. 1046–1054 .
[144] A. Pozdnoukhov , C. Kaiser , Space-time dynamics of topics in streaming text,in: Proceedings of the Third ACM SIGSPATIAL International Workshop on Lo-
cation-based Social Networks, ACM, 2011, pp. 1–8 . [145] Z. Yin , L. Cao , J. Han , C. Zhai , T. Huang , Geographical topic discovery and com-
parison, in: Proceedings of the Twentieth International Conference on World
Wide Web, ACM, 2011, pp. 247–256 . [146] V.W. Zheng , Y. Zheng , X. Xie , Q. Yang , Collaborative location and activity rec-
ommendations with GPS history data, in: Proceedings of the Nineteenth In-ternational Conference on World Wide Web, ACM, 2010, pp. 1029–1038 .
[147] P. Symeonidis , A. Papadimitriou , Y. Manolopoulos , P. Senkul , I. Toroslu , Geo–social recommendations based on incremental tensor reduction and local
path traversal, in: Proceedings of the Third ACM SIGSPATIAL International
Workshop on Location-Based Social Networks, ACM, 2011, pp. 89–96 . [148] H. Kawakubo , K. Yanai , GeoVisualRank: a ranking method of geotagged im-
agesconsidering visual similarity and geo-location proximity, in: Proceedingsof the Twentieth International Conference Companion on World Wide Web,
ACM, 2011, pp. 69–70 .
[149] A. Silva , B. Martins , Tag recommendation for georeferenced photos, in: Pro-ceedings of the Third ACM SIGSPATIAL International Workshop on Loca-
tion-Based Social Networks, ACM, 2011, pp. 57–64 . [150] T. Sandholm , H. Ung , Real-time, location-aware collaborative filtering of web
content, in: Proceedings of the 2011 Workshop on Context-Awareness in Re-trieval and Recommendation, ACM, 2011, pp. 14–18 .
[151] R. Pálovics , P. Szalai , J. Pap , E. Frigó, L. Kocsis , A .A . Benczúr , Location-awareonline learning for top-k recommendation, Pervasive Mob. Comput. 38 (2017)
490–504 .
[152] R. Pálovics , A .A . Benczúr , L. Kocsis , T. Kiss , E. Frigó, Exploiting temporal influ-ence in online recommendation, in: Proceedings of the Eighth ACM Confer-
ence on Recommender Systems, ACM, 2014, pp. 273–280 . [153] M. Harvey , F. Crestani , Long time, no tweets! time-aware personalised hash-
tag suggestion, in: Proceedings of the 2015 European Conference on Informa-tion Retrieval, Springer, 2015, pp. 581–592 .
[154] S. Androutsellis-Theotokis , D. Spinellis , A survey of peer-to-peer content dis-
tribution technologies, ACM Comput. Surv. (CSUR) 36 (4) (2004) 335–371 . [155] P. Han , B. Xie , F. Yang , R. Shen , A scalable P2P recommender system based on
distributed collaborative filtering, Expert Syst. Appl. 27 (2) (2004) 203–210 . [156] B.N. Miller , J.A. Konstan , J. Riedl , PocketLens: toward a personal recommender
system, ACM Trans. Inf. Syst. (TOIS) 22 (3) (2004) 437–476 . [157] J.K. Kim , H.K. Kim , Y.H. Cho , A user-oriented contents recommendation sys-
tem in peer-to-peer architecture, Expert Syst. Appl. 34 (1) (2008) 300–
312 . [158] Y. Zhao , J. Ye , T. Henderson , The effect of privacy concerns on privacy rec-
ommenders, in: Proceedings of the Twenty-first International Conference onIntelligent User Interfaces, ACM, 2016, pp. 218–227 .
[159] S.K.T. Lam, D. Frankowski, J. Riedl, Do You Trust Your Recommendations? AnExploration of Security and Privacy Issues in Recommender Systems, Springer
Berlin Heidelberg, Berlin, Heidelberg, pp. 14–29. 10.1007/11766155_2
[160] A. De Spindler , M.C. Norrie , M. Grossniklaus , Collaborative filtering based onopportunistic information sharing in mobile ad-hoc networks, in: Proceedings
of the 2007 OTM Confederated International Conferences” On the Move toMeaningful Internet Systems”, Springer, 2007, pp. 408–416 .
[161] R. Schifanella , A. Panisson , C. Gena , G. Ruffo , MobHinter: epidemic collabora-tive filtering and self-organization in mobile ad-hoc networks, in: Proceed-
ings of the 2008 ACM Conference on Recommender Systems, ACM, 2008,
pp. 27–34 . [162] D. Karamshuk , C. Boldrini , M. Conti , A. Passarella , Human mobility models for
opportunistic networks, IEEE Commun. Mag. 49 (12) (2011) 157–165 . [163] L. Del Prete , L. Capra , Differs: a mobile recommender service, in: Proceedings
of the 2010 Eleventh International Conference on Mobile Data Management(MDM), IEEE, 2010, pp. 21–26 .
[164] N. Eagle, A.S. Pentland, MIT Reality mining dataset, 2005, (Downloaded from
https://crawdad.org/mit/reality/20050701 ). doi: 10.15783/C71S31 . [165] Y. Zhao , J. Ye , T. Henderson , A robust reputation-based location-privacy
recommender system using opportunistic networks, in: Proceedings of theEighth EAI International Conference on Mobile Computing, Applications and
Services, ACM, 2016 . [166] A.R. Beresford , F. Stajano , Location privacy in pervasive computing, IEEE Per-
vasive Comput. 2 (1) (2003) 46–55 . [167] S.K. Lam , J. Riedl , Shilling recommender systems for fun and profit, in: Pro-
ceedings of the Thirteenth International Conference on World Wide Web,
ACM, 2004, pp. 393–402 . [168] R. Burke , M.P. O’Mahony , N.J. Hurley , Robust collaborative recommendation,
in: Recommender Systems Handbook, Springer, 2015, pp. 961–995 . [169] I. Parris, T. Henderson, F.B. Abdesslem, St_andrews/locshare dataset,
2011, (Downloaded from https://crawdad.org/st _ andrews/locshare/20111012 ).doi: 10.15783/C7WW2F .
[170] G. Lo Giusto , A.J. Mashhadi , L. Capra , Folksonomy-based reasoning for content
dissemination in mobile settings, in: Proceedings of the Fifth ACM Workshopon Challenged Networks, ACM, 2010, pp. 39–46 .
[171] A.J. Mashhadi , S.B. Mokhtar , L. Capra , Habit: leveraging human mobility andsocial network for efficient content dissemination in delay tolerant networks,
in: Proceedings of the 172 IEEE International Symposium on a World of Wire-less, Mobile and Multimedia Networks & Workshops, WoWMoM 2009, IEEE,
2009, pp. 1–6 .
[172] C. Boldrini , A. Passarella , HCMM: modelling spatial and temporal propertiesof human mobility driven by users’ social relationships, Comput. Commun.
33 (9) (2010) 1056–1074 . [173] M. Salathé, M. Kazandjieva , J.W. Lee , P. Levis , M.W. Feldman , J.H. Jones , A
high-resolution human contact network for infectious disease transmission,Proc. Natl. Acad. Sci. 107 (51) (2010) 22020–22025 .
[174] A. Keränen , J. Ott , T. Kärkkäinen , The one simulator for DTN protocol evalu-
ation, in: Proceedings of the Second International Conference on SimulationTools and Techniques, ICST (Institute for Computer Sciences, Social-Informat-
ics and Telecommunications Engineering), 2009, p. 55 . [175] F. Ekman , A. Keränen , J. Karvo , J. Ott , Working day movement model, in: Pro-
ceedings of the First ACM SIGMOBILE Workshop on Mobility Models, ACM,2008, pp. 33–40 .
[176] M. Pasinato , C.E. Mello , M.-A. Aufaure , G. Zimbrao , Generating synthetic data
for context-aware recommender systems, in: Proceedings of the 2013 BRICSCongress on Computational Intelligence and Eleventh Brazilian Congress on
Computational Intelligence (BRICS-CCI & CBIC), IEEE, 2013, pp. 563–567 .
M.G. Campana, F. Delmastro / Online Social Networks and Media 3–4 (2017) 75–97 97
P
a
t
t
P
[177] M. del Carmen Rodríguez-Hernández , S. Ilarri , R. Hermoso , R. Trillo-Lado ,DataGenCARS: a generator of synthetic data for the evaluation of con-
text-aware recommendation systems, Pervasive Mob. Comput. 38 (2017)516–541 .
[178] G. Guo , J. Zhang , Z. Sun , N. Yorke-Smith , LibRec: a java library for recom-mender systems., in: Proceedings of the 2015 Conference on User Modeling,
Adaptation, and Personalization (UMAP Workshops), 2015 . [179] M.D. Ekstrand , M. Ludwig , J.A. Konstan , J.T. Riedl , Rethinking the recom-
mender research ecosystem: reproducibility, openness, and LensKit, in: Pro-
ceedings of the Fifth ACM Conference on Recommender Systems, ACM, 2011,pp. 133–140 .
[180] P. Castells , N.J. Hurley , S. Vargas , Novelty and diversity in recommender sys-tems, in: Recommender Systems Handbook, Springer, 2015, pp. 881–918 .
[181] Y. Zheng , B. Mobasher , R. Burke , CarsKit: a java-based context-aware recom-mendation engine, in: Proceedings of the 2015 IEEE International Conference
on Data Mining Workshop (ICDMW), IEEE, 2015, pp. 1668–1671 .
Mattia Giovanni Campana is a Ph.D. student in ComputerEngineering at the University of Pisa in collaboration with
the Institute for Informatics and Telematics of the Na-tional Research Council of Italy (CNR). He received a M.S.
degree in Computer Science from the University of Milanin 2015. His research activity is related to the study and
application of Social Recommender Systems to optimize
the data dissemination and content discovery in Oppor-tunistic Networks and Mobile Social Networks.
Franca Delmastro is a researcher of the IIT Institute of
the National Research Council of Italy (CNR). Her researchinterests are in the field of Future Internet and Cyber-
Physical systems, working on communication protocols
for wireless and opportunistic networks, mobile middle-ware solutions, recommender systems for online and mo-
bile social networks, context and social-aware mobile ap-plications and personalized m-health solutions. She pub-
lished several works on international journals, confer-ences and book chapters in the pervasive and mobile
computing area. She participated in the organization of
several international conferences. She served as TPC ViceChair of IEEE PerCom 2017, Workshops co-Chair of IEEE
erCom 2014, General Co-Chair of Work-in-Progress session of IEEE PerCom 2012,nd General Chair of some international workshops. She is currently member of
he Editorial Board of International Journal of Communications, Network and Sys-em Sciences (IJCNS) and she has been Guest Editor of Special Issues for Elsevier
ervasive and Mobile Computing Journal e Computer Communication Journal.